4
MODEL COMPONENT
Business Intelligence allows decision makers to have a better understanding of the context
of their choices. It is based upon the collection and examination of information called
"analytics." Analytics are the result of some kind of modeling of (usually) historical data
that generally includes the application of statistical analysis, operations research, or other
quantitative tool for the purpose of either explaining what is or predicting what will be.
The purpose of the model is to represent critical relationships in such a way to guide
decision makers toward a desired goal. The involvement
and support of
these models
is what differentiates DSS from other kinds of computerized systems. Said differently,
without a model, a system is not a DSS. Hence, to understand DSS, one must understand
models. Unfortunately, in practice, modeling, and especially model management, is the
least developed of the aspects of DSS.
MODELS AND ANALYTICS
Modeling is the simplification of some phenomenon for the purpose of understanding its
behavior. Even before the tsunami of data began hitting organizations, modeling provided
a structure for understanding and predicting events. Modeling simplifies and abstracts
detailed event data to allow understanding of the major forces acting upon the alternatives.
It involves a process of summarizing and accumulating of data. In addition, modeling
involves a process of removing unnecessary detail, thereby allowing the important patterns
to shine through. This is similar to what is illustrated in Figure
4.1.
All of the panels have
Decision Support Systems for Business Intelligence
by Vicki L. Sauter
Copyright © 2010 John Wiley & Sons, Inc.






126
MODEL COMPONENT
Figure 4.1. Process of modeling.
the word "model" in the middle. In the upper panels, you cannot discern the word because
there is too much detail—the important factors are not put together and the unimportant
factors act as noise clouding the image. Slowly, as some factors are accumulated and
the irrelevant noise is removed from the panel (as you move down), it is possible to see
increasing amounts of the word until at the bottom the word is perfectly clear. So it is
with modeling. The key is to identify patterns in the data; one must identify the critical
components and scrape away the others until the important trends are apparent. As you
can see in this diagram, you begin with too much detail to identify any patterns. Once you
scrape away some detail, you begin to see a variety of issues that are not of interest to
your decision situation. Scraping those away, you find a variety of distractions to the core
purpose of your modeling. As those distractions are eliminated, you begin to model. Before






MODELS AND ANALYTICS
127
Figure 4.2. A model airplane.
you get too much clarity, you must first eliminate the unimportant variables. Then finally,
once all of those other issues and the remaining small details are gone, it is possible to see
the model clearly. That is how the modeling process works.
Most people have their first experience with models as children, such as in model
airplane building or model trains. Everyone knows that a model airplane is not a real
airplane and hence will not perform all the functions of a real airplane. However, certain
attributes of the plane are created realistically, such as the number of
wings,
the number of
propellers, the relative size or colors of the plane, and its markings (Figure 4.2). That is,
model makers do not include all of the details of the plane but rather only those that are
important to understand whatever aspects of the plane that are important to the decision
maker. A child might be able to ascertain the development of planes by noting the evolution
of number and placement of wings, the use of propellers, and even how the shape of
the plane has changed over time. Another child, with different interests, might use these
models to learn the colors and markings of planes associated with different countries.
Hence, the amount of detail and the kind of detail necessary for the model airplanes are
dependent upon the interests of the child at that moment. In other words, whether or not
the model is sufficient is dependent upon the needs of the decision maker (in this case,
the child).
Business modeling fulfills the same objective. The purpose of a model is to simplify
the choice context so that decision makers can understand options and their ramifica-
tions clearly. When statisticians develop regression models, their goal is to determine the
factors essential to understanding the variability in the phenomenon of interest. Market
research specialists, for instance, use regression to predict demand for a particular product.
They understand that many factors affect a person's decision whether or not to purchase a
product. However, in developing their marketing campaigns, it is useful to know whether
their product appeals to young, unmarried professionals or to retired blue collar work-
ers,
and whether the desirability of the product is different in different regions of the
country.
Most business decisions have a large number of influential factors, and decision makers
need to filter the essential components of the situation from the irrelevant ones. While it
seems obvious that models fill this need, not everyone feels comfortable with models. It
may not be clear what model is most appropriate. Other times it is clear what kind of
model is needed, but the data are not there to support it. Finally, some decision makers may






128
MODEL COMPONENT
not know how to interpret the results, especially if that means understanding the model's
sensitivity to particular market conditions.
Although models can be applied without DSS, their power is magnified with DSS
because of the inherent flexibility, friendly interfaces, and query capability of
DSS.
Histor-
ically, decision makers needed to rely upon others to develop and interpret models for them
because of the difficulty of running the computer programs associated with models. With
DSS,
decision makers are given personal access to appropriate models and appropriate data
and immediate access to results.
It is this
easy and friendly
access that makes DSS-based models so attractive. Decision
makers can understand the implications of their judgment and modify those judgments
when they appear to be inconsistent with what is known. In addition, because of the speed
and efficiency of analysis, decision makers can examine more alternatives so as to find
a good strategy. Furthermore, the model encourages decision makers to investigate the
variables that are most sensitive to assumptions. Improvement in these aspects of problem
AIDS PLAN Is a DSS resource that allows health care workers in Great Britain to plan re-
sources for HIV/A1DS-related services better The system explicitly encourages decision makers
to focus on
^what-if"
questions so they can creatively experiment with strategies that might
prove useful in meeting the needs of this increasing care-needing group. The DSS can be
used to explore the consequences of alternative strategies or investment
s in resources as well
as the sensitivity of those consequences to particular assumptions about uncontrollable and
unpredictable factors
, This in turn allows decision makers to examine the impacts of the de-
cisions in terms of likely overload, need for further resources, and flexibility to meet future
uncertainties.
Forecasts of demand within particular localities are derived from the COX National Forecasts
by patient categories. Decision makers can elect whether to examine these forecasts at their
low, medium, or high range. This projection of patient demand in turn forms the basis for
experimentatio
n with care options, Costs-of-care options by patient category are used to estimate
the costs and resources required to treat the projected patient demand,
The model's analysis is based on a division of patients into categories that, for planning
purposes, can be considered relatively homogeneous in their demand for services, Criteria that
can be used to classify patients include clinical state, possible drug abuse, age, dependency
,
housing situation, and the presence or absence of informal support at home.
For each category, the health authority needs to identify alternative care options, A care
option is a costed combination of service inputs that constitutes a clinically acceptable method of
treating or supporting a member of the client group. It is defined in terms of the basic resources
needed to supply appropriate care and treatment. Model users can adopt the list of resources
provided with AIDSPLAN or change it to suit their special concerns or circumstances
. Up to 32
different resources can be accommodated in the model, Once users have established such lists of
resources, they can express any given care option as a particular combination of recourses from
the list in specified amounts.
For any particular assumptions made about future demand, A IDS PLAN computes the re-
sources and cost consequences of the identified care strategy. Using a menu, the user can display
summaries of the results at different levels to see the effect of the input assumptions and
to
identify
where further analyses may be needed. In fact, medical personnel currently are using AIDSPLAN
to facilitate discussion of the consequences for services of using AZT prophylactically and the
impact of day care facilities on the provision of inpatient beds.






OPTIONS FOR MODELS
129
analysis in turn aids decision makers in advocacy and implementation of the chosen solution
because they understand more facets of the problem better. For example, the New Zealand
yacht-racing team exploited the benefits of alternative generation and evaluation in its
design of Black Magic 1 and 2, which competed in the America's Cup in 1995. Over
10,000 options were considered during the four-month competition, which allowed the
team to make constant improvements in the design of the yachts at the waterfront facility.
Many believe this systematic evaluation of alternatives led to the remarkable performance
in which the New Zealand team swept the field 5 to 0.
OPTIONS FOR MODELS
A model is a generalized description of a decision environment. The goal of creating it
is to simplify a phenomenon in order to understand its behavior. While that is a nice
definition, it does not help decision makers to understand how to model or even to identify
a model.
Decision support systems can include several types of models, some of which you
have studied in your other classes. For example, statistical models include regression
analyses, analysis of variance, and exponential smoothing. Accounting models include
depreciation methods, budgets, tax
plans,
and cost
analysis.
Personnel models might include
"in-basket" simulations or role playing. Marketing models include advertising strategy
analyses, consumer choice models, and product switch models. The characteristics of these
models differ substantially, as do their uses; each represents simplification of a decision
phenomenon that is useful for understanding some component of behavior. The skills
needed to build and use these models and the kinds of support needed to help less skillful
users utilize the models effectively also differ considerably. Part of the challenge of creating
a DSS is knowing what models need to be included and how they can be supplemented to
make them meaningful and useful for the decision maker.
To determine what kind of model to use, generally we need two kinds of information:
what the decision maker needs and the kinds of data available to use. Since models are
simplifications of real situations that act as vehicles for learning about those situations, we
need to select a model that helps to answer the questions that decision makers pose. Also,
since models have underlying assumptions about the data that are used, we can only select
models for which the assumptions are appropriate for the available data. We will use a
variety of dimensions to describe models and the role they fulfill in decision making, as
shown in Table 4.1.
Table 4.1. Dimensionality
of Models
Representation
Time Dimension
Linearity of
the
Relationship
Deterministic
vs.
Stochastic
Descriptive
vs.
Normative
Causality
vs.
Correlation
Methodology Dimension






MODEL COMPONENT
Representation
The first dimension, the representation, describes the kind of data needed in a model which,
in turn, dictates the necessary approaches used to collect and process the data. In particular,
we are distinguishing between models that rely upon experiential data and those that rely
upon objective data. The difference between the two is the process by which the model is
generated, not the answer that is derived.
Experiential models rely upon the preparation and information processing of peo-
ple,
either individually or as a group. These models might include judgments, expert
opinions, and subjective estimates. For example, diagnostic software used by physi-
cians to help in prescribing treatment for tumors or blood diseases models the expe-
rience of expert practitioners. Similarly, a forensic animation simulation was created
to convict a Florida man of vehicular homicide. The simulation showed how his truck
drove into a group of children (one of whom was killed) and then left the scene of the
accident.
One of
the
problems associated with the use of such models is their subjectivity in use.
In such modeling, the information used and the manner in which it is used to make a choice
are up to the decision maker. If two individuals attempt use the same behavioral model, they
may come to different conclusions because they are drawing upon different experiences
and are likely to weight those experiences differently. In the case of
the
forensic simulation,
the verdict was appealed on the basis of the use of the simulation which, according to
the defense, misrepresented the scene of the accident (which happened at night) and the
automobile.
Objective models, on the other hand, rely upon specified, detached data and its analysis
by known techniques. They are considered "objective" because the data considered and
the way they are used are specified, constant, and independent of the specific decision
maker's experiences. Consider the Advanced Trading System from Scottrade shown in
Figure
4.3.
This system allows decision makers to access real-time stock quotes, historical
data, and models for analyzing the data. The return on investment computed by one user
for a particular option will be the same as the return on investment computed by another
user for that same option. Hence, there is no subjectivity associated with the analysis.
However, that in no way means that they are unbiased or lead everyone to the same
conclusion. Clearly, we can bias results by the selection of the variable, time period, or sam-
ple group. For example, conclusions about the yield of investments can vary substantially
by the time horizon considered; stock market investments tend to provide poor yields when
examined over short time horizons but excellent yields, on average, when examined over
multiple decades. Both provide an "objective" view of the performance of a portfolio, yet
they provide very different conclusions; not providing both views presents a biased view
of the problem. The ability to recognize such biases and thereby study multiple aspects of
a problem is one of the advantages of using a DSS.
Neither the experiential nor the objective model is appropriate all the
time,
and each has
its own strengths and weaknesses. Objective models have the advantage of being straightfor-
ward to apply and easily replicated with new data. In addition, they can save time in that they
do not require the establishment of extensive experience such as is needed for some forms of
behavioral modeling. These models have limitations as well. The basic assumption underly-
ing objective modeling is that the simplification of reality necessary to create
a
mathematical
model does not eliminate the essential issues controlling the decision environment. That
is,
it
assumes that
the
most important factors, such
as
competition, regulation,
prices,
and technol-
ogy, are represented in the simplified model in a manner similar to that in the actual decision






OPTIONS FOR MODELS
131
Figure 4.3. Screenshot from Scottrade's Advanced Trading System Software. The image is reprinted here with the
permission of Scottrade.
environment. If these factors change in a significant
way,
the mathematical models would not
be appropriate because the essence of the decision environment and its probable reactions
would not be represented. Under these circumstances, it is important to rely on experiential
models.
Some DSS allow for the integration of both the objective and experiential models. For
example, the DSS facilitating the U.S. Army plans for future needs of materiel incorpo-
rates both kinds of modeling. Objective models are built based upon quantitative analysis
of historical data. In this case, the historical data represent past demands for and uses of
the materiel over time. The projections combine models that first assume a continuation
of past patters of materiel use and then take into account planned activities such as ma-
jor exercises. These forecasts are supplemented with heuristics about possible changes in
the needs during the upcoming time horizon; expert opinions and human judgment are
included to alter the projections. The DSS helps the user evaluate the combined model per-
formance by continually measuring trends and alerting the decision maker to changes in the
trends.






132
MODEL COMPONENT
Time Dimension
The time dimension identifies how much of the activity of the decision environment is
being considered. The two ends of the continuum are static models and dynamic models.
At the static end, models represent a snapshot in time of all factors affecting the decision
environment. Such models assume that everything will remain the same. Similarly, such
models assume that there is no dependence of later decisions or actions on the choice under
consideration. Dynamic models, on the other hand, consider the decision environment over
some specified time period. They may consider the same phenomenon during different
periods of time or interrelated decisions that will be considered during different time
periods.
Time can be represented in models in a variety of
ways.
An example of
its
explicit use
of time to examine some phenomenon is shown in Chapter 2. The software in use at Gap
Minder (example results of which are shown in Figure 4.4) looks at how the factors under
consideration change by increments of
a
year. In this way decision makers can examine the
Design Insights
Modeling Chip Architectur
e
Designing chip architecture for the best performance and smallest size is an exceedingly difficult
task. Today, computers solve the problem by considering possible combinations. They are fast, but
the computer lacks both intuition and visual pattern recognition
. These are not only characteristic
s
at which humans excel but also characteristics that could yield a better or even optimal design.
Researchers at the University of Michigan are developing mechanisms to combine the speed of
computers and the skill of humans in a project called FunSAX By solving problems using the
Fun SAT board, players contribut
e to the design of complex computer systems. Although the
humans believe they are just selecting actions that will turn all buttons green, they are in fact
solving complex problems or selecting the best arrangemen
t of options. The solution is then given
to a computer scientist who translates that solution into hardware design. The researchers hope
to use this combination of objective and subjective modeling to improve chip designs, databases,
and even robotics. Perhaps someday similar
l
'games" can be used to improve other decisions,
Adapted from
Dc
Orio,
A.
and
V.
Bertac^o
,
Design Automation Conference
(DAC),
San Francesco,
CA,
avail
able
at;
http ://w w w,
eecs.
um ich. edu/—va per ia/rcscarch/publicaüons/DAC09Fun STA .pdf July
2009.
Used
with
permission of Mr
r
De
Οτίο and
Dr,
Bertacco.
The Fun SAT
1+
game"
is
available
at
http://funsaLeecs.unieh.edu.






OPTIONS FOR MODELS
133
Figure 4.4. Looking at intervals of time for patterns. Gapminder software shows incremental
changes in the graph on an annual basis. In this way, decision makers can examine the relation-
ship's changing nature over time. Visualization from Gapminder World, powered by Trendalyzer
from www.gapminder.org.






134
MODEL COMPONENT
directionality of the change, the times at which the magnitude of change shifted direction,
and the relative change of
a
variety of observations. Other ways of representing time include
using time as a variable in the model, examination of results in a "before" and "after" time
period, and using models that use interdependence of time periods explicitly, such as with
dynamic programming.
Linearity of the Relationship
This third factor of a model is called linearity. It refers to the relationship between two or
more factors. Such relationships are either linear or nonlinear. Everyone has seen a linear
relationship in two dimensions; it is expressed by a straight line. It can be interpreted easily
as the more of
x,
the more of v. For example, the larger the warehouse, the greater the
storage volume available.
Anything other than the straight line is referred to as a nonlinear relationship. The
two-dimensional graphs in Figure 4.5 and the three-dimensional graph in Figure 4.6 are
nonlinear. Nonlinear relationships require the user to specify the kind of relationship be-
tween and among the variables. For example, sales related to the natural log of advertising
expenditures, or sales related to the square root of price, or sales related to the square of
time spent with a sales representative are all nonlinear relationships. Such relationships
do not have the nice intuitive interpretation of linear models. Nor is it obvious how to
build the model. The linearity, or lack
thereof,
affects the kind of model that one can use.
For most linear model solution techniques there are parallel nonlinear solution techniques.
The nonlinear models are more complex. The first—and hardest step of the process—is
to specify the nature of the relationship. While it may be easy to determine the kind of
relationships shown in Figure 4.5, if you have some mathematical background, data rarely
Figure 4.5. Nonlinear relationships.






OPTIONS FOR MODELS
135
Figure 4.6. Nonlinear higher dimension relationship.
come so well behaved. Generally they include error terms such as that shown in Figure 4.7.
Clearly this makes it harder to determine exactly what nonlinear function should be used
with the data. It takes time, patience, experience,
and
an understanding of the phenomenon
being modeled to get it right.
For this reason, it is tempting just to use the linear model to approximate the nonlinear
data. Not only does that avoid the problem of having to determine the underlying function,
but also the linear models are better behaved, easier and faster to solve, and generally have a
straightforward approach to solution. There are times when such approximations are good
enough, especially since the techniques of nonlinear models generally are harder and slower
to solve, require a hierarchical approach, and often result in "good answers" rather than
"the best answers." Other times, however, the conclusions gained from the linear model are
inappropriate for the nonlinear world.
Deterministic Versus Stochastic
Most of the modeling taught in business colleges is deterministic. For example, consider
linear regression. You might want to predict sales using price and advertising. To do this,
you collect past data about the three variables and run the regression. You might result with
something like
Sales,
= 5.64 + 16.1 Price, + 0.58 Advertising,






136
MODEL COMPONENT
Figure 4.7. Nonlinearity with randomness.
To use the model, a decision maker would substitute in values for price and advertising
and from it would get an expected value for sales. This is deterministic in that it uses fixed
variables that are determined by averaging the error terms over the training data set. It is a
method that works well for many situations.
Stochastic modeling, on the other hand, explicitly uses probabilistic distributions for
one or more variables in the model to view how situations might evolve over time. These
models use historical data, but rather than the specific variable associations (as shown with
the regression above), stochastic models use the fluctuations in the data
to
determine a likely
underlying probability distribution for one or more of
the
variables. Using those underlying
distributions, a model is constructed to reflect the scenarios, decision points, and outside
influences on the system. The model then is run hundreds or thousands of
times
so decision
makers can view the range of the impact as well as specific estimates. The most common
form of stochastic modeling is based on Monte Carlo analysis. As you can see from Figure
4.8,
the result of such an analysis is the outcome of some variable, say production rates,
associated with each different combination of randomly generated parameters. We then
look for the average or typical value (the middle line) and for the typical range of values
(between the outer
lines).
The questions we might answer is whether these values are "good
enough" or perhaps how we can improve the values by changing things (such as adding
another line in the production facility).
Descriptive Versus Normative
Another choice to make selecting a model is whether you wish it to be descriptive or
normative. Descriptive models are those that report what is happening in the data. It might
be the sales of widgets by division, the profitability of a sales
line,
the absenteeism associated






OPTIONS FOR MODELS
137
Figure 4.8. Results of a Monte Carlo analysis.
with
a
particular facility, or the number of radio advertisements run by the competitor. These
descriptive models might be created to provide decision makers with a quantitative view
of what is happening in the organization or part of an organization as background or for
monitoring. Or, the descriptive models might serve as predictive analytics, which attempt
to forecast how factors such as sales, profitability, absenteeism, or competitor's ads will
behave in the future. As said earlier, of course, such models are only valid if the factos
pushing on the phenomena are the same in the future as they have been in the past.
The alternative to descriptive models is normative models. These models represent
an
ideal value
of sales, profitability, or absenteeism in an organization. The output of the
normative models, perhaps of sales, is then compared to the actual sales to determine if
operations are running as we expect they should. This form of modeling does not provide
a view of how the organization is changing, that is, how sales are growing (or declining)
over time, but rather simply a view of how the current organization is competing relative
to a set of standards or values.
Causality Versus Correlation
The relationship between correlation and causality is one of the most misunderstood and
misapplied in all of modeling. Correlation, however it is represented, refers to the amount
and direction that two or more variables vary together. It might be thought of
as
the level at
which the variables simultaneously change. If two variables move positively together, that






MODEL COMPONENT
means as one increases, so does the other; if the correlation is negative, the variables move
in opposite directions. Similarly, the magnitude of the correlation indicates how similar
the movement is a larger correlation means the rate of change of the two variables is more
similar.
However, correlation does
not
say anything about what caused this association. For
example, there is a positive correlation between education and income. The fact that
education and income are correlated does not imply that getting more education causes
your income to increase. It is possible that people from higher incomes simply get more
education. Or, it is possible that there is another factor, say intelligence, that causes changes
in both variables. Similarly, increasing price does not cause a drop in sales. It may be that
reduced sales causes a company to increase its price to cover its costs. Or, it may mean that
a competitor is pushing both variables to change. If the goal in the analysis is to determine
what
causes
changes in some factor, then in addition to correlation, it is necessary to prove
that it is impossible that anything else but one factor could have caused the change in the
other factor. This requires the design of a scientific experiment that controls the variables
to approximate such a counterfactual state of the world. Generally this is achieved by
conducting experiments on identical items or randomizing exposure to the experimental
factors.
Methodology Dimension
The last dimension, methodology, addresses how the data (whether objective or experien-
tial) will be collected and processed. There are five general methodologies: (a) complete
enumeration, (b) algorithmic, (c) heuristic, (d) simulations, and (e) analytical. In complete
enumeration, by far the hardest and most expensive option, information about
all
feasible
options is collected and evaluated. Under many circumstances, complete enumeration
is totally impractical. However, there are some contexts for which it is necessary or
desirable. For example, the U.S. Census is an example of complete enumeration in which
all individuals in the United States are identified and
counted}
The purpose of counting
all individuals is to understand the population shifts in the United States so representation
in the Congress can reflect actual population density. Rather than sampling various areas
in each state, the government identifies every person individually.
Complete enumeration also has been useful in the application of neural networks
of transaction files for pattern recognition. For example, a neural network system was
constructed for Mellon Bank of Chicago to identify suspicious credit card activity that
might be indicative of stolen credit cards. Historically, both human auditors and electronic
expert systems identified dubious transactions through abrupt increases in either the number
or the size of transactions. By examining all the transactions, the neural network identified
a change in
small
purchases as an indicator of stolen credit cards. In fact, at that time, card
thieves were using small purchases, often as little as $1, in pay-at-the-pump gas stations,
to determine whether the cards were still being accepted. It was this complete enumeration
of transactions, supplemented by pattern recognition capabilities, that allowed the system
to respond quickly to the presence of criminal behavior.
The second approach, the algorithmic model, is the development of a set of proce-
dures that can be repeated and will, eventually, define the desired characteristics of the
1
It
has been noted
that
the
U.S.
Census process does
not count
homeless
individuals and underestimates
their
numbers.
Strictly speaking, then, the census is not
a
complete enumeration.






OPTIONS FOR MODELS
decision environment. Such models are best represented by the field of operations re-
search/management science. Algorithms have a set of repetitive calculations that can be
implemented to find the best answer. The set of calculations itself
is
based upon the charac-
teristics of
a
particular
problem.
Unlike total enumeration, an algorithm identifies promising
information that can be used to identify the best outcome without first evaluating all pos-
sible options. An example of such a modeling technique is the Simplex Algorithm. To use
this model, we need to represent a problem as a linear program, determining an objective
function that can be optimized (either maximized or minimized) and a set of constraints.
Typically the objective function uses the minimization of
costs,
the maximization of utility,
or some related concept. The constraints define the availability of scarce resources such
as time, money, and inputs. If we can represent the problem as a linear program, we can
use repetitive operations based upon matrix row reduction calculations and find the best
solution to the problem.
2
These repetitive operations are simple arithmetic operations; the
process of applying them is the algorithm.
Algorithms are used widely today in business, organizations, and government. They
can help decision makers know how to place investments, where to advertise products, or
how to assign staff to projects. One area where algorithms are used heavily is in personnel
planning and scheduling. For example, many hospital systems use algorithms to assign
nurses and other staff to shifts. In some cases, the systems include measures of "intensity"
of patient illnesses so that they can determine whether the optimal general staffing levels
will meet the specific needs on a daily basis. Similarly, the U.S. Army uses an algorithm-
based DSS called
ELIM-COMPLIP
with input from other modeling forecasting systems
to plan for deployment of personnel to various tasks so as to meet their strength needs as
specified in the Force Structure Allowance.
The third possible model process is heuristic. Generally heuristics are applied to large
or ill-structured problems that cannot be solved algorithmically. The goal is to find a
satisfactory solution that is reasonably close to optimal. All heuristics involve searching,
evaluating, learning, and more searching to find a good solution. They are usually developed
for a particular problem in order to take advantage of the structure of a problem. Some
heuristics are designed to construct solutions; others are designed to improve existing
solutions. Since heuristics are so dependent upon a particular representation of a problem,
they are not often generalizable to other problems.
Heuristics can be quantitative solutions to a problem or behavioral solutions to a
problem. In the former
case,
the model is a numeric representation of
a
choice and we focus
on numeric processing. Typically, a quantitative heuristic is developed as an alternative
to using a quantitative algorithmic approach, if, for example, a reliable algorithm is not
available, if
the
computation time is excessive, if
the
data are limited, or if
the
problem is so
big it cannot be reasonably simplified otherwise. For example, if the decision variables in
a problem are restricted to dichotomous (0-1) values or integer values, known algorithms
may fail to find an optimal solution. This might include a firm's assignment of production
processes
to
particular production facilities or
a
financial
institution's assignment of deposits
to lockboxes. Similarly, if the objective to a problem is nonlinear, or if there are many
variables or constraints, known algorithms may fail to find an optimal solution. Some
heuristics can be identified that take advantage of the mathematical structure of a problem
to find good answers to these problems.
2
There are some special problem structures that cannot be solved using this algorithm. In addition,
some problems cannot be solved
practically
with this technique because the number of variables
and/or constraints is so large it would take
a
prohibitively long amount of
time
to solve the problem.






140 MODEL COMPONENT
Modeling Insights
Linear Programming
To understand algorithms and their use, let us consider a specific problem. An MIS Club plans to
sell two special fruit baskets for the upcoming holiday season. Fruit basket A contains 3 apples, 4
oranges, and I honeydew melon and sells for
$8,
Fruit basket
B
contains 4 apples, 3 oranges, and
2 honeydew melons and sells for $12. The amount of each fruit available and their costs to the
MIS Club are shown in the table below. If it is assume
d that the MIS Club can sell all the baskets
it makes, how many of each one should they make?
Quantity
Available
Cost per
Piece
Apple
Orange
Melon
160
180
60
$030
$0.20
$L20
The first step is to represent the problem mathematically
. In this case, we will have two
variables, x and y, where
x
represent
s the number of fruit basket A to make and
y
represents the
number of fruit basket B to make. We know that each fruit basket A sells for $8 and each fruit
basket B sells for $12, but in order to know how much profit wc will make, wc must compute the
costs of each basket. Basket A contains 3 apples at
S
,30,4 oranges at $ .20, and I melon at SI ,20,
so it costs $2.90 to make up the basket (if we assume the actual basket is free)· Hence, the net
profit from basket A is $5.10. Using a similar method, we can find that the net profit from Basket
B is $7.80. Hence, our objective is to:
Maximize 5-10X +7.80y
However, there are constraints dictating the availability of fruits which must be met. Using the
quantities above, they are:
Apples
Oranges
Melons
3x-h4>< 160
Ax
+ 3y < 180
\x-\-2y
<60
Conceptually, the algorithm for solving this problem looks at possible values for
x
and
y
and
selects the one that maximizes our objective. Consider the graph below;
80
- Oranges constraint
Apples constraint
Melon constraint
40 60 80
Number of fruit basket A






OPTIONS FOR MODELS
141
If the heuristic is behavioral, then we consider the relationships between concepts and
use symbolic processing of the data. In fact, this kind of behavioral heuristic is generally
referred to as expert systems (a branch of artificial intelligence). Expert systems use rules,
frames, objects, and metarules (often referred to as demons
3
) to replicate the solution
3
The term "demon" in
a
programming environment refers to
a
portion of
code
that lies dormant until
a particular event, such as the change in the value of a variable, causes the code to process. These
demons might cause particular actions to occur, such as the searching of a database, or they might
prohibit actions to occur and to take the user along a different path of
code.
The algorithm
"knows"
to look for the feasible combinations of the two types of fruit baskets,
as shaded in the-graph. Further it "knows" that the best combination is going to be one of the
four "extreme" or comer points highlighted above. The algorithm evaluate
s an extreme point
with regard to the objective
(5.1 Ox
-h 7.80;y). It then looks at the adjacent corners to determin
e
if one of them give a better solution. If so, the algorithm moves to that new point and begins
again. In essence, the algorithm moves from comer to corner, always improving the value of the
objective. With large problems, the process is important because one can have many variables and
many constraints resulting in millions of corner points. Since the algorithm follows a systematic
approach to improvement, it ends up checking only a small percentage of the possible points. In
this case, it is the combination of 36 fruit baskets of type A and 12 fruit baskets of type B, giving
a profit of $277
+
20 to the MIS Club.
DSS in Action
MLB Schedules as Models
Baseball is called the "great American pasttime" because so many Americans share a passion
for the game. The game may live or die by the pitcher or the next power batter, but the schedule
is dependen
t on modeling, The Sports Scheduling Group (556) uses mathematical programming
and high-performance computers running virtually nonstop for months to develop a schedule for
major league baseball. According to one of the partners of SSG, "a typical model for a sports
scheduling problem is a combinatoria
l design with nasty side constraints and multi-objectives.'*
Schedule makers deal with conflicting requirement
s and preferences as a matter of course,
but as the financial and competitive stakes in athletics rise, so does the complexity of creating a
balanced schedule
. To maximize revenue, it is crucial to have important games televised on the
right days and times. These requirements frequently conflict with more traditional requirement
s
of a 'Tair" schedule that balances strength of schedule, home and away games, and travel.
SSG must consider the following constraints when developing a schedule:
* Each club plays 162 games and 52 series, including 13 at home on weekends.
* Games within each month and during summer dates should be reasonably balanced
between teams.
* Single-series and four-series home stands and road trips should be minimized; two- and
three-series home stands and road trips are preferred.
* No more than four series home stands or road trips should be scheduled,
* There should be no doubleheaders in the original schedule
.
* Considerations must be made to the miles traveled by one team during a season. No team
should travel in excess of 50,000 miles over the course of the season,
* Three game series are optimal (minimize number of
two-
or four-game series),
In addition, SSG entertains the requests of the teams, the television network
s broadcasting the
games, and the MLB Players Union*






142
MODEL COMPONENT
technique that an expert would use to solve an ill-structured, nonquantifiable problem.
These models can give meaning and context to the symbol and incorporate subjective
information about the validity of an answer or the way in which the answer should be used
to obtain a solution.
The fourth approach to modeling is simulation. Unlike algorithmic and heuristic mod-
eling, which provide a normative answer, simulation provides descriptive results. The
goal of simulation is to imitate reality either quantitatively or behaviorally. Typically, this
Modeling Insights
Presidential Selection Heuristic
s
Every four years there is a great deal of money spent on trying to predict who will win the
U.S.
presidentia
l election. Pundits examine the various segments of the population carefully and
determine the issues that are most important for each group, who best addresses those issues (for
the groups)
, and what the likelihood of that group voting will be. There arc millions of dollars
spent to predict who is likely to win the election
* As the viewing public knows, there are many
flaws to these predictions.
Allan J. Lie htm an, professor of history at The American University in Washington, D.C.,
looks at the situation in a different way. He applied statistical pattern recognition algorithm from
seismology to the question of who would be elected. Professor Lichtman began with nearly 200
questions, which were all binary (yes-or-no) variables, and the algorithm picked those which
displayed the greatest difference between the proportion of the time the variable was "yes"
for years when the incumbent party won and the corresponding proportion for years when the
challenging party won using all
U*S*
elections starting with 1860 as the training set. Over lime,
he narrowed it to
13
keys. They are:
1.
The incumbent party holds more seats in the ILS, House of Representative
s after the
midterm election than after the preceding midterm election.
2
+
There is no serious contest for the incumbent-part
y nomination.
3,
The incumbent-part
y candidate is the current president
4,
There is no significant third-party or independent candidacy.
5,
The economy is not in recession during the campaign.
6, Real (constant-dollar) pcr-capita economic growth during the term equals or exceeds
mean growth for the precedin
g two terms.
7,
The administratio
n has effected major policy changes during the term.
8, There has been no major social unrest during the term,
9, The incumben
t administratio
n is untainted by major scandal.
10.
There has been no major military or foreign-pol icy failure during the term.
11.
There has been a major military or foreign-policy success during the lerm.
12.
The incumbent is charismatic or is a national hero.
13.
The challenger is not charismatic and is not a national hero.
According to Dr. Liehtman's models, if six or more of these statements are false, the incumbent
party loses the popular vote. Using that criterion, the model has only been wrong twice, in 1876
and 1888. Of course, in the United States, it is the electoral vote, not die popular vote, that
determines the winner, so sometimes this method does not predict who will actually be in the
White House.
Samulson, D. "Road to the White House;'
ORMS
Today,
Vol 35, No 5, October 2008. This material is
reprinted with permission of
the
publisher and the author.






OPTIONS FOR MODELS 143
Negotiation Ninjas, developed by researchers at Southampto
n University, are intelligent agents
that use heuristics to help bring together buyers and sellers on the shopping website Aroxo, The
agents use a series of simple rules—known as heuristics—to find the optimal price tor both
buyer and seller. The heuristic
s guide not only the price but also the ways to address multiple
simultaneous negotiations
. Sellers must answer a series of questions about how much of
a
discount
they are prepared to offer, whether they are prepared to go lower after a certain number of sales
or at a certain time of day, and how eager they are to make a sale. Buyers only need to identify
the item they wish to purchase and the price they are willing to pay for it, The agents then act
as an intermediary, scouring the lists of sellers who are programmed to accept a price in the
region of the one offered
. If they find a match, the seller is prompted to automatically reply with
a personalized offer. The buyer then has a choice to accept, reject, or negotiate. If they choose to
negotiate, the agent analyzes the seller's criteria to see if they can make a better offer. The process
continues until either there is a sale or one of the parties pulls out.
One system using nonquantitative heuristic
s is
PROSPECTOR.
The purpose of this system is to
predict mineral deposits given geological information about a region. Some of
PROSPECTORS
rules arc the following,
* RULE 1: IF the igneous rocks in the region have a fine to medium grain size, THEN they
have a porphyritic texture (0,5)«
* RULE
2:
IF the igneous rocks in the region have a fine to medium grain size, THEN they
have a texture suggestive of a hypabyssal regional environment (2,
(λΟΟΟΟΟ
I).
* RULE 3; IF the igneous rocks in the region have a fine to medium grain size and they
have a porphyriti
c texture, THEN they have a texture suggestive of a hypabyssal regional
environment (100, 0.0000001).
* RULE 4: IF the igneous rocks in the region have a texture suggestive of a hypabyssal
regional environment, THEN the region is a hypabyssal regional environment (65, 0.01).
* RULE 5: IF the igneous rocks in the region have a morphology suggestive of a hy-
pabyssal regional environment
, THEN the region is a hypabyssal regional environment
(300,ΟΌ001),
* RULE 6: IF the region is a hypabyssal regional environment, THEN the region has a
favorable level of erosion (200, 0.0002).
* RULE 7; IF Coeval volcanic rocks are present in the region, THEN the region has a
favorable level of erosion (800, I).
The system processes these and other rules much the way an expert geologist would to examine
the geological, geophysical, and geochemical, data to predict where ore-grade minerals could be
found. The numbers in parentheses indicate measures of certainty with the conclusions that are
built into the reasoning process.
Source:
Waterman, D. A. (1986)
A Guide to Expert Systems,
"Prospector Rules,
TT
p. 58. Reproduced with
permission of Pearson Education. Inc.






144
MODEL COMPONENT
Figure 4.9. Simulation with animation. (Source: The Great Flu, Erasmus University, available: http://thegreatflu.
com.) Application was developed jointly by Erasmus University Medical Center and the Ranj. Serious Games. Image is
reproduced with permission.
involves the repetition of an experiment and the description of the characteristics of certain
variables over time. For example, a simulation of a factory would include a variable that
measures the amount of time an average part spends waiting in lines and the amount of
time it takes to process the inventory. Using the mathematics underlying the simulation, we
could vary the demand for products, the raw material arrivals, and the number and types
of production lines and study the impact of these variations on the amount of time one part
spends waiting in line and making a transaction. With today's simulation software, decision
makers can vary decision variables and see the impact with animation.
Consider the simulation shown in Figure 4.9. This simulation was created by influenza
researchers at Erasmus Medical University Center in Rotterdam to help decision makers
examine the activities associated with fighting a pandemic influenza outbreak. Once the
simulation has started, decision makers have a variety of actions they can take in each region
ranging from improving research facilities, to stockpiling vaccines and antiviral medicines,
to isolating sick individuals; closing schools, markets, and airports; or simply stating
warnings. Each activity costs money, and the decision maker is given a budget. During
the simulation, decision makers can view information about the spread of the virus across
the world and the number of resulting deaths. Through use of such simulation, decision
makers can experiment with various strategies and gauge their effectiveness without putting
a single person in jeopardy of the illness.
Simulations help decision makers understand how external influences can affect the
outcome of their decision. For example, computer companies rely heavily upon simulation
in deciding when to introduce new models. Simulations model customer demand, pricing,
and dealer inventories and simulate a variety of relevant conditions, such as component






OPTIONS FOR MODELS
145
price changes or even the impact of a rival model. In this way, the managers can evaluate
the risk
before
taking the risk.
Similarly, personnel departments use "in-basket" simulation exercises to help indi-
vidual managers determine the best approaches to addressing the problems that arise in
managing people. In this case, the manager measures not a mathematical variable, but
The U.S. military is one of the most significant users of simulations in the world today. The
Generalized Air Mobility
Model,
or GAMM, simulates the entire theater airlift system's movement
of cargo from source to destination. Hence, the DSS provides simulation of flights, airdrops,
overland cargo transshipment, and survivability of cargo in the various modes of transportation
.
(The DSS does
not
simulate the outcome of the campaign, just the ability of the airlift system to
meet the operational demands of a given scenario.)
The quality of the insight from this simulation, as in any simulation, comes from the quality
of the measures that were built into the system for evaluation. Historically, the military used
measures such as rate-of-cargo movement, average aircraft flying time per day, utilization rate,
and departure reliability While these measures provide some indication of the basic throughput
of the operation, they do not measure the effectiveness of the mission or how it supports combat
forces. Hence, GAMM has factors of evaluation such as:
• Timeliness of deliveries
• Effectiveness in making multifligh
t deliveries within narrow time and location constraints
such as those necessary for combat mission
s
• Ability to move large, oversize items
In addition to providing operational logistics for a particular campaign, GAMM also can
predict where long-term airlift characteristics need to be changed and hence offer insights into
fum re designs.
The costs of providing health care have skyrocketed over the last 20 years. At
the
same time the
incidence of infections, especially antibiotic-resistant infections, contracted during hospitalization
has increased significantly
. States have recognized the impact of these secondary (not existing
upon admission) infection
s on health cane costs, and some have introduced legislation to reduce
payments to hospitals with high rates of secondary infection. Clearly it is in everyone's best
interest to reduce the incidence of infections contracted during hospitalization
. But, this is a
difficult problem to solve due to interactions among the various pathogens, categories of illness
of the patients, and occupancy rate of the hospital. In other words, it is hard to know "where to
start." However, researchers worked with Cook County Hospital in Chicago to build simulations to
represent various scenarios of these variables so they could study the relative efficacy of improve
d
hand-hygiene protocols versus changes in patient isolation policies. They found both policies
could have a significant impact on the rate of infections. However, when they also examined
the costs—both to the patient and to the hospital—under various conditions, they determined
that improved hand-hygiene protocols were more appropriate as a first approach to solving the
problem. Further, the researchers provided insights into condition
s where the policies should be
changed and what the associated costs would be.






146
MODEL COMPONENT
rather the reaction of another individual in order to experiment with more positive and more
negative reactions and determine which will provide the desired effect. Finally, today's
technology can make it possible to simulate how it feels to drive a given automobile over a
variety of surfaces and in a variety of conditions to determine which car provides the most
desirable ride given its cost.
The essence of constructing simulation models is to simplify the elementary relation-
ships and interdependencies of the situation being considered. While it does simplify the
conditions, simulation also allows us to build in real-life complexities that might affect the
variables being measured. It is descriptive in its answer, thereby encouraging
"what-if"
kinds of experimentation in which many alternatives can be considered independently, and
time is compressed so that long-term effects can be measured quickly.
Design Insights
Modeling Failures
Computer simulations are not replicas of reality. For example, Boeing Co. Engineers used simula-
tion to design a fuse pin that held the engines to the wing for its 747 cargo plane. After El Al Israel
Airlines had a crash in 1992, where the plane killed over 40 people in the Netherlands, engineers
reviewed their simulation. They found that the simulation had missed several weak points in the
design of the fuse pin. The fuse pin had in fact broken, causing the crash.
odeling Insights
Finding bin Laden
Professors in the Geography Department at UCLA applied biogeographic models to the question
of locating Osama bin Laden in the spring of
2009.
Biogeographi
c models use known properties
of plants and animals to predict how they will distribute themselves over space and time. These
models were applied to publicly available satellite imagery.
The particular models employed are called a
"
'distance
decay theory" and "island biogeog-
raphy theory/' They were employed because they are associated with the distribution of life and
extinction. Distance decay theory states that as one goes further away from a precise location,
there is an exponential decline in the turnover of species and a lower probability of finding the
same composition of
species.
The theory of island biogeography states that large and close islands
will have higher immigration rates and support more species with lower extinction rates than small
isolated islands,
These theories can be applied over varying spatial scales to posit bin Laden \s location based
on his last reputed geographic location. Distance decay theory would predict that he is closest to
the point where he was last reported and, by extension, within a region that has a similar physical
environment and cultural composition (that is, similar religious and political beliefs). For instance,
the further he moves from his last reported location into the more secular parts of Pakistan or
into India, the greater the probability that he will find himself in different cultural surroundings,
thereby increasing the probabilit
y of his being captured or eliminated. Island biogeographic theory
predicts that bin Laden is in a larger town rather than a smaller and more isolated town where the
extinction rate would be higher. Finally, high-resolutio
n analyses of a city can be undertaken to
identify individual buildings that match bin Laden's life history characteristics. For example, he
reportedl
y has a small entourage of body guards, requiring a structure that contains at least three
rooms.
Using these methods, the biogeographers identified not only a specific town in Pakistan in
which bin Laden is likely to be located but also three specific buildings in which he is likely to be
located. However, no national security agency has commented on whether they have applied this
methodology or whether or not the professors were accurate.






PROBLEMS OF MODELS
Simulations are not without their disadvantages, however. They do not provide an
optimal solution; instead they provide information about conditions from which we can
glean a good or possibly optimal solution. Like heuristics, inferences are not transferable
beyond the specific type of problem being considered. Finally, and most important, the
construction of simulations can be slow and costly.
The last type of methodology is the analytical model. Analytical modeling refers to
the process of breaking up a whole into its parts and the associated process of examin-
ing the parts to determine their nature, proportion, function, and interrelationships. Where
phenomena are well defined, analytical approaches solve for related variables that have
specified properties within limits. For example, the phenomenon of gravity is well de-
fined so that we can use specified equations to describe how an object will fall. Where
phenomena are not well defined, which includes virtually all business-related phenomena,
the analytical approach determines how to separate a given problem into its constituent
parts and determine what subcomponents are most important in affecting the interactions
with other subcomponents. Statistical analyses, especially regression and other predictive
models, provide good examples of analytical modeling.
Consider, for example, the process of creating strategies for football games. The
interdependence of the players and the complexity of the plays make it difficult for any
individual to make choices without help. National Football League teams use DSS with
sophisticated analyses to make these decisions. The DSS helps the coach to understand the
tendencies of
his
own team and the opposition and hence to plan strategies that will respond
to them. The New England Patriots use a DSS to select the best players at the lowest cost to
decide what play to run or whether to challenge a referee's ruling and even how to improve
total fan experience.
PROBLEMS OF MODELS
Modeling is not without its problems. Modeling depends on understanding the factors that
impact the phenomenon of interest and using those variables in the correct proportion. The
failure to identify an important variable, to select an inappropriate time horizon, or
to
overfit
the model to some time period will decrease the value of the model to the decision makers.
Quantitative modeling, in addition, assumes that the factors acting upon the phenomenon
will continue to be important and will continue to work in the same fashion as in the past.
For example, most public transportation companies have models to predict ridership. They
use the models to decide routes for buses and trains and how often to schedule vehicles
on each route. If done well, the models provide a good mechanism for planning. However,
when gasoline prices suddenly surge, the assumptions about ridership change significantly
and the models no longer provide a reliable output for decision making. The use of models
assumes the underlying assumptions continue to be true. Decision makers need to consider
if that is true.
Not knowing if
the
assumptions are true is one problem. Knowing the assumptions are
not true and continuing to use the models make their use more hazardous. Consider the
financial institutions and their use of models prior to the recession of
2009.
Analysts chose
to program their risk management systems with overly optimistic assumptions and to feed
them oversimplified data. In other words, financial analysts modeled the system so as not
to identify all of the risks and perhaps maybe even the correct risks. Rather than noting
recent volatility in the market, the models looked at several years of trading history, which
dampened the impact of an impending crisis. Others, it is claimed, developed models that
did not reflect the complexity of the financial products being traded.






MODEL COMPONENT
DATA MINING
One kind of modeling that is particularly important in DSS is data mining. When we think
of mining, we think about digging deeply into some repository to find something of value.
When one mines for diamonds, one digs into seemingly common rocks to find brilliant
pieces of carbon. Said differently, one needs to look carefully through vast repositories of
useless rock to find that one nugget that is valuable. A similar process is used for data
mining. Data mining might easily be defined as the process of extracting valuable patterns
from a mass of data. Companies often mine the data to find evidence of theft or fraud,
patterns of purchasing (or other behavior of interest), or evidence of the need for new
products, new markets, or new sources of
revenue.
This is not a new idea; companies have
been trying to mine their data for hundreds of years. What
is
new is that companies are
able to collect and save much more data now than ever before. Similarly, although there
have been many data-mining tools available for some time, today's processing power has
brought us new tools that increase our ability to find patterns in the data.
Consider, for example, one of the largest procurers of data in the world today, Google.
Every day, there are several million searches on Google to find anything from a product
for a gift to health information. Google saves the data. It is not just the search, but if you
have logged in, it saves your name and email address, the date and time of day, and your
Internet Protocol (IP) address. The IP address, of course, gives Google information about
how you are connected to the Internet and the country (and, in the U.S., the city and state)
from which you connected to the Internet. There have been hundreds of billions of searches
since 1997, when the search engine was launched.
Google mines its data, meaning that it attempts to find patterns in the searches that are
useful. One such mining exercise is the Google attempt to predict influenza outbreaks. It
compared the number of queries about influenza with traditional flu surveillance systems,
such as the CDC process in the United States. Google tracked the searches for appropriate
terms by geographic area in the United States between 2003 and 2008 and compared it
to publically available data from the CDC's U.S. Influenza Sentinel Provider Surveillance
Network. Google researches found not only that the search results were verified by the CDC
data but also that the search results
predicited
the CDC data. That
is,
because people search
for symptoms prior to seeking a physician's care (from which the CDC data are compiled)
and because it is so much faster to process the search data than the physician's data, Google
could predict the outbreaks
by region
up to two weeks earlier than the CDC. The data are
shown in Figure 4.10.
4
If
the
data continue to provide the same predictive capabilities, they
could predict pandemics or epidemics sooner and thus give health professionals a longer
window to stem the negative effects.
A famous example of data mining is from a chain of midwestern (U.S.) grocery stores
and the purchasing data of their customers. They found male customers generally shopped
on Thursdays and Saturdays. Further they found that these men tended to do their weekly
shopping on Saturdays, but only purchased a few items on Thursdays. Further analysis of
what
they purchased showed that men who purchased diapers on Thursdays also tended to
purchase beer. Armed with this result, the grocery chain made sure the beer display was
close to the diaper display
and
that both diapers and beer were sold at full price on Thursday
to maximize their revenues.
4
More information about Google's work with flu trends can be found at http://www.google.org/
flutrends/. Information about their other data-mining
activities
can
be
found at http://www.google.org/.






DATA MINING
149
Figure 4.10. Google results. Adapted from Google's Flu analysis, available at http://www.
google.org/flutrends.
Of course, not all data mining is good. Consider the research conducted by Acquisti
and Gross (2009) of Carnegie Mellon University. They showed that it is possible to predict
narrow ranges in which a given person's social security number is likely to fall simply by
mining public data. In particular, they used the Social Security Death Master File, which
includes information about those people whose deaths have been reported to the Social
Security Administration. Since this file is a popular tool among genealogy researchers
for finding ancestors, it is readily available online and easy to search. The researchers
mined the data to detect statistical patterns in Social Security Number (SSN) assignment
associated with date of birth (and thus likely date of application for a SSN) and location of
birth. Using their results, they were able to identify the first five digits of
44%
of deceased
individuals born in the United States from 1989 to 2003 and
complete
SSNs with less
than a thousand attempts for almost 10% of those deceased individuals. With that tool,
it becomes statistically likely that they could predict with the same level of accuracy for
living individuals. The professors are interested in the mining algorithms and the public
policy implications; however, in the wrong hands, this could provide the keys needed for
identity theft.
Although data warehouses provide access to information that will help decision makers
understand their operations and environment better, users can become lost in the enormous
possibilities for analysis and miss the forest for the trees. These efforts require the co-
ordinated efforts of various experts, stakeholders, or departments throughout an entire
organization. Available tool users mine the value of the information available in these
warehouses to find the kinds of data that seem to discriminate among alternatives the best,
identify cases which meet some criterion, and then summarize the result or find patterns in
the data to highlight important trends or actionable situations.
The five approaches to data modeling are given in Table 4.2. In each case, the goal is
to find patterns in the data that we might exploit to improve the business. Knowing what
items customers tend to purchase together, or under what conditions emergency rooms will
need assistance, or when products are sufficiently similar to substitute them, will all help
managers run their businesses better. It requires that the system search for patterns in the
data and then differentiate the patterns that are interesting and useful from those that are
illusions and spurious. Said differently, the goal is to find a model that generates predictions






150
MODEL COMPONENT
Table 4.2. Data-Mining Goals
Classifications
Clusters
Regressions
Sequences
Forecasting
that are most similar to the data on which you build the model. At the same time, however,
the focus is not on the training data, but rather on future data. If you overfit to your training
data, then the patterns are likely to perform less well on test set data. Said differently, it
provides a model that is specific to the random fluctuations in the original data. When
applied (which is always the goal), over-fit models tend to do poorly because the new data
experience different "random fluctuations." Hence it is important to have "pure" (not used
in the original analysis) data on which to test any mining model before using it to impact
business rules. Netflix understood that, but apparently some of the contestants did not.
The most commonly used data-mining technique is
classification.
Classification iden-
tifies patterns that are unique for members of a particular group. It examines existing items
that have already been classified and infers a set of rules from them. For example, the
system might classify attributes of those students who complete their degree in a specified
number of years from those who do not. By finding the similarities among students who
do not successfully complete their degrees, the system might find "early warning signals"
on which the administration can act.
Classification mining can produce misleading results, particularly if not done properly.
For example, one of the most controversial classification efforts was the Total Information
Awareness Program (ITAP) of the U.S. Department of Defense. The original goal of the
Nctfiix is known for using quantitative analyses for improving its performance. In 2006 it an-
nounced a
S1
million competition to the first team that could improve it recommendation system
by 10%. The recommendatio
n system, which is used to suggest movies to individual customers,
predicts whether someone will enjoy a movie based on how much tbey liked or disliked other
movies. Netflix provided anonymous rating data for mining and a test data set to evaluate how
closely predicted ratings ot movies match subsequent actual ratings. This set off a flurry of activity
of individuals
, groups, and groups of groups, fn mid-2009
:
a team called BellKor's Pragmatic
Chaos was the first to achieve the goal of improvin
g the system by 10Ό9%
+
According to the
rules,
the other teams had 30 days to improve upon BellKor's method. Just before the deadlin
e
was reached another team, The Ensemble, submitted a method that improved the rating system
by 10,10%, BcllKor did not have time to respond.
However, shortly thereafter
, the team's captain, Yehuda Korcn posted a note on his blog that
he was contacted by Netflix and was told they have the best test accuracy and should be declared
the winner. Why? It appears that Netflix kept two verification test sets: one that was the basis
for the public standings and another that was secret The winner was selected based on the success
of the approach on the
secrer
data set. So BellKor, which appeared to come in second, based
on the public verification test set, seems poised to be the winner based on the hidden test set
Apparently The Ensemble got their additional improvement by overfilling their algorithm to the
test data set; when tested on the unused data, their algorithm was inferior.






DATA MINING
program was to examine large quantities of data, from telephone calls and credit card
purchases to travel and financial data, to detect data that would identify potential terrorists.
TIAP was to use both supervised and unsupervised learning to identify "people of
interest." Supervised learning might find rules linking certain fields in the databases with
known terrorist behavior. Using this method, the mining algorithm might identify all indi-
viduals from certain countries who enrolled in flight school but did not learn how to land
and see what else they had in common. Examination of the additional fields might help
decision makers identify those having terrorist intentions. Unsupervised learning might
find people engaged in suspicious activities that are not necessarily terrorist oriented but
are unusual and should be investigated.
The program was quickly canceled because of the concern about constitutionality of
abuse of the privacy rights of
U.S.
citizens associated with the program. But, if it were not
cancelled, could it work?
This project highlights some of the difficulties of data mining.
False
Positives.
In practice, any time you try to classify people, some will be incorrectly
classified. Some people who should, using this example, be classified as terrorists
would not be (called a false negative). Further, some who should not be classified
would be classified as terrorists; that is a false positive. Even rules that were 99%
accurate (and that level of accuracy would
be
phenomenally unlikely) would identify
a substantial number of false positives. Consider that when looking at 200 million
individuals a
1 %
error rate still generates 2 million false positives. That would result
in not only possible negative impacts on a large number of lives but also a lot of
wasted investigation time.
Insufficient
Training
Sets.
Fortunately, there have only been a small number of instances
of
terrorism.
With such small data sets, the resulting rules would be far less accurate
than the 99% identified in the previous point.
Pattern Changes.
Following this approach, all analyses are done on historical data.
Any behavior changes in the terrorists over time would not be represented.
Anomalies.
People sometimes change their behavior for perfectly good reasons having
nothing to do with terrorism. So, even though they may fit a "profile" for a terrorist
(or for a fraudulent charge), it may have nothing to do with terrorism.
Because the costs of being wrong are so high in this situation and because of the constitu-
tional issues, the program was stopped. But these same issues can impact any data-mining
situation and need to be addressed before decisions are contemplated.
A similar process is
clustering.
The process identifies clusters of observations that
are similar to one another and infers rules about groups that differentiate them from other
groups. It differs from classification, however, in that there are no items a priori classified,
and hence the model needs to determine the groupings as well. A university might cluster
students of similar performance in a class for the purpose of the studying what pastclasses or
experiences they share that might explain their similar performance. Credit card companies
regularly cluster records to determine which customers are likely to respond to different
incentives or even which charges are likely to be fraudulent.
A third kind of data mining is known as
regression.
The goal of this kind of data
mining is to search for relationships among variables and find a model which predicts
those relationships with the least error. For example, a supermarket might gather data of
what each customer buys. Using association rule learning, the supermarket can work out






152 MODEL COMPONENT
Modeling Insights
nderstandin
g single-malt Scotch Whiskey
Single-malt Scotch whiskeys are an acquired taste. They are distilled from barley at a single
distillery and matured in oak casks for at least three years (some for many years). Scotch whiskeys
cannot be matured in new oak casks because the new oak would overpower the taste of the whiskey,
so it is only matured in used casts, Clearly the previous use of the cask will impact the taste of
the Scotch whiskeys. The taste of American bourbon in oak will impact the taste differently than
will Portugese port or by Spanish sherry or Carribean rum or maderia. Similarly, each year that
the Scotch whiskey is in the cask will change the taste since it continues to process. The water
supply will also impact the taste of the final product.
Single-malt Scotch whiskeys tend to be categorized by the region in which they were
produced, While this is useful for those who really know their whiskey, it is less useful for the
general public. So,
a
project called Whisky Classified develope
d
a
clustering system to help people
understand styles of the common brands. Said differently, the project helps someone answer the
question, ^if I Like this brand, what other brands am I likely to like?"
The developers reviewed tasting notes in recently published books on malt whiskey and
from distilleries, From this, they developed a vocabulary of 50U aromatic and taste descrip-
tors for Scotch whiskey. They applied these terms to 86 single-mall Scotch whiskey using
a product called ClustanGraphics. The cluster analysis groups malts into the same cluster
when they have broadly the same taste characteristics across all 12 sensory variables. Tech-
nically, the method minimizes the variance within clusters and maximizes the variance between
clusters.
The result was 10 clusters of single-malt Scotch whiskeys:
Cluster
A:
Full-Bodied
, Medium-Sweet, Pronounced Sherry with Fruity, Spicy, Malty Notes
and Nutty, Smoky Hints
Cluster B: Medium-Bodied, Medium-Sweet, with Nutty, Malty, Floral, Honey and Fruity
Notes
Cluster C; Medium-Bodied, Medium-Sweet, with Fruity, Floral, Honey, Malty Notes and
Spicy Hints
Cluster D: Light, Medium-Sweet, Low or No Peat, with Fruity, Floral, Malty Notes and
Nutty Hints
Cluster E: Light, Medium-Sweet, Low Peat, with Floral, Malty Notes and Fruity, Spicy,
Honey Hints
Cluster F: Medium-Bodied, Medium-Sweet, Low Peat, Malty Notes and Sherry, Honey,
Spicy Hints
Cluster G: Medium-Bodied, Sweet, Low Peat and Floral Notes
Cluster H: Medium-Bodied, Medium-Sweet, with Smoky, Fruity, Spicy Notes and Floral,
Nutty Hints
Cluster I; Medium-Light, Dry, with Smoky, Spicy, Honey Notes and Nutty, Floral Hints
Cluster
J;
Full-Bodied, Dry, Pungent, Peaty and Medicinal, with Spicy, Feinty Notes
Those who want more information about the exercise and especially advice about other Scotch
whiskeys they might enjoy should consult Wishart (20O6).
Adapted from
Wishart,
D.,
Whiskey
Classified*
London,
Pavillion,
2006*
Materials used with
the
permissio
n
of
Mr. Wishart and Pavillion, and imprint of Anova Books.






DATA MINING
what products are frequently bought together, which is useful for marketing purposes. This
is sometimes referred to as "market basket analysis." One uses this kind of mining to
find associations among the factors. Associations are events linked with regard to a single
criterion, such as two or more courses that students tend to take together, such as DSS and
database systems. The fact that students take the courses together might not be apparent
without the analysis. However, after the analysis, we know that the two courses should not
be scheduled at the same time.
Sequences
are events linked over some period of time, such as patterns the students
employ for taking courses over multiple semesters. The important characteristic of these
linkages is that they are ordered: observations with characteristic X are also likely to have
characteristic
Y.
For
example,
a student who takes a statistics course this semester
is
unlikely
to take the forecasting course for two subsequent semesters. This will help the department
plan course offerings. Or, perhaps more commonly, voters who express interest in issues
of education and health care prior to the election are more likely to vote for the Democratic
candidate.
Finally,
forecasting or predictive
data mining is the process of estimating the future
value of some variable. While clearly it is possible to use tools like regression to forecast
the future, the goal of the data mining is to find rules that might predict what will happen.
Universities do (or should do) this kind of mining since they have significant historical
databases of students, their characteristics prior to admission, and their level of success.
So a data-mining exercise might identify specific combinations of test scores, experience,
and grades that were associated with successful students (generally defined as those who
graduate) to find decision rules for admissions. Insurance companies mine their data of
symptoms, illnesses, and treatment plans and outcomes to determine the best course of
treatment for particular illnesses. In the latter case, this analysis might be with regard to
outcome and to cost.
An interesting form of predictive data mining is in the area of text mining. This can
be particularly useful for brainstorming or alternative generation in the decision-making
process. Suppose, for example, that you are the state senator on a transportation committee
and you are trying to determine what projects are most important to your constituents. Of
course you can read everything on the Internet about it or you can poll your constituents, but
both of those take time. Instead you want to have the computer analyze some transportation
blogs on the subject of transportation in your state. One way to analyze the blog is to input
the text of the blogs in a
a
product such as IBM's "Many Eyes" so it can analyze the words in
the
text.
A starting point might
be
to examine
a
word cloud such
as
that shown in Figure
4.11.
The word cloud sizes the words in proportion to the number of times that they appeared
in the blog. You can see from this that the bloggers discuss specific locations, such as St.
Louis or Jefferson City, and individuals in the MoDOT hierarchy most, because those words
are the largest. Moving beyond this, you see that terms such as "bridges," "safety belts,"
and "work zones" appear frequently. To pursue those lines further, consider a word tree that
gives more information of how those words are used in context. An example that shows
phrases following the word "bridge" is shown in Figure 4.12. Many Eyes will allow users
to click on the various terms and follow them to their completion or do additional analyses
on them. The goal of the use is, of course, to provide ideas to the senator about what is
important to his or her constituency.
Data mining can be a very useful tool for identifying trends that decision makers might
not have considered. However, it can also identify statistically significant trends that are not
the least bit useful. The decision maker needs to understand the assumptions underlying the
statistics and the implications for their data before applying the results from data mining.






Figure 4.11. Word cloud analysis of a blog discussing plans and problems of projects under consideration by Missouri DoT Department of Transportation.
The summary was prepared by David Doom using IBM's
tool,
"Many Eyes."






Figure 4.12. Word tree of a blog. Word bridge analysis of a blog discussing plans and problems under consideratio
n by Missouri DoT. The summary was
prepared by David Doom using IBM's
tool,
"Many Eyes."






156
MODEL COMPONENT
For example, large sample sizes can result in even very small differences to be statistically
significant. Even if you know that a rare event is statistically more likely under certain
circumstances, it might not change how you approach a decision. If it does not change the
decision, it is not important. Also, statistical significance does not address the question of
the cost of gaining and using the intelligence. If the cost of applying the rule is greater than
the savings associated with ignoring it, even if it is statistically significant, the exercise
does not imply that decisions should be changed. Finally, there is the problem of running
many tests. If you test enough hypotheses, about 5% of them should be "significant" even if
they are all false. That is what the significance level means. So, if the tests are random, and
not based upon some reasonable understanding of the business, some results might simply
reflect spurious relationships that are not useful for running the business.
A variety of analytical tools—neural networks, decision trees, rule induction, and data
visualization—as well as conventional analyses are used to complete these
five
kinds of data
mining. These tools can "learn" to predict changes in the environment, generate rules for
classification of
data,
find similar subjects among the data, identify if-then rules for action,
and display data so that decision makers can glean important
patterns.
To be successful, the
approach and the product must meet the needs of the user and a particular data warehouse.
Other criteria for the evaluation of data-mining products are listed in Table 4.2.
Intelligent Agents
Intelligent agents are pieces of software that complete specific, repetitive tasks on behalf
of
the user. They are not new to computers; in fact, they are commonly in use on systems to
monitor CPU and peripheral use and capacity. Other intelligent agents are associated with
e-mail
systems, where they help sort and prioritize
e-mail
by sender or topic on behalf of
the user. Their new use is as a means to search through relational databases to find relevant
data for decision makers. Even more exciting is the combination of search protocols with
analytical capabilities that will cause the intelligent agent not only to find data but also to
analyze it to find examples of trends or patterns the decision maker might miss on his or
her own. In addition, the intelligent agent can get at the information faster to detect unusual
occurrences so the decision maker can act upon them more quickly.
For example, consider the product DSS Agent from MicroStrategy, as shown in Figure
4.13.
This product surfs a data warehouse for information and summarizes it for decision
makers. In fact, this particular screen summarizes data at Recovery.gov, which in 2009 con-
taied information about how the American Recovery and Reinvestment Act was working,
including an up-to-date data on the expenditure of funds.
Researchers at the University of Vermont developed a website, http://www.weteelfine.org4
s
that
mines through some 2.3 million blogs looking for sentences beginning with
tl
I feel" or "I am
feeling." They use personal online writing to determine the mood of people in
real
rime.
After
mining the sentence, they use the standardized "psychologica
l valence" of words (established by
the Affective Norms for English Words) to give each sentence a happiness score, The rating of
the individual blog is not important; rather their goal is lo measure the big picture of a town or
other grouping of
people.
They use their tool in an exploratory fashion to measure the feelings of
the country as a whole. Clearly such a tool could be used to mine for other words, such as those
of a company's product, to provide decision makers with consumer's attitudes about the product,






DATA MINING 157
Satoshi Kanazawa, a reader in management and research methodology at the London School of
Economics, published a series of papers that predict the sex of one's baby, the last of which
is "Beautiful Parents Have More Daughters"*, Dr, Kanazawa took a sample of almost 3000
individuals who were asked the number of children of each gender and who were rated on a
five-point scale regarding attractiveness. His results are shown in the following graph as the
points.
60%
-, 1
Two researchers reexamined his method and found that the "statistical significance
" noted
in the original paper just did not exist.f
Note that the least attractive people (rated 1) had about a 50-50 chance of having a girl
while the most attractive people (rated 5) had about a 56% chance of having a girl. What the
author did was to compare the aggregate of groups 1-4 to group 5 and found that the difference
between them was significant. But, in reality, a correct statistical test would have made not only
that comparison but also other combinations of groups, such as group I to the aggregate of groups
2-5,
or the aggregate of groups
1
and 2 to the aggregate of groups 3-5, and so on. Furthermore, if
you do those additional tests, they
must
be included in the test of significanc
e of the experiment,
In other words, statistical validity relies not just upon the one comparison but rather on
all
of
the comparisons
together.
As the authors point out, the curved lines in the diagram above are the
result of a better test; this test does
not
show statistical significance. This is one of the examples
of statistical problems associated with the mining of data.
*From S. Kanazawa, "Beautiful Parents Have More Daughters: A Further Implication of the Generalize
d
Trivers-Willard
Hypothesis,'*
Journal
of
Theoretical
Biology,
244,
2007,
pp.
133-140.
r
From A. Gelman and D. Weakllem, Of Beauty, Sex and Power,"
American
Scientist,
97(4), July-August
2009,
pp.
310-314,
I






158
MODEL COMPONENT
Figure 4.13. DSSAgent screen. Summary of the progress of the American Recovery and Reinvstment Act available
at: http://www.microstrategy.com/recovery-act-data
. Image is used with permission from microstrategy.
Using these intelligent agents, users can schedule intelligent agents to execute on
a one-time basis, periodically, or based upon events. For example, decision makers can
perform regular scanning of absenteeism or missing reports to highlight indicators that
problems might need attention. Or, decision makers can schedule intelligent agents to
find information about changes in demand after planned promotions or after a particular
indicator reaches some prespecified value. Using workflow triggers, users can specify both
pre-
and postagent macros that can integrate with other modeling components of the DSS.
For example, the agent could find information that would automatically be imported to a
forecasting application to compute projected demand. If desired, another agent could be
triggered to mail results of the application automatically to people on the management team.
Many intelligent agents today provide a set of options through which the user can scan
the data warehouse. For example, users can define filters based upon specific qualifying
criteria. Or, users can define percentile and rank filtering; using this option, decision makers
could identify the source of the top 10% of their raw materials, for example. Similarly,
intelligent agents can be launched using conditional metrics. Hence, users can specify
information to be found regarding a particular business unit and compared it to that of
multiple business units or to the company as a whole.
To fully exploit the data-mining capability, however, the intelligent agents need to be
combined with artificial intelligence so the software can find not only the data but also






MODEL-BASED MANAGEMENTSYSTEMS
159
the patterns in the data. In fact, if it works well, data mining should find answers in the
data which the decision maker has yet to consider asking. Data-mining tools find patterns
in the data, infer rules from them, and then refine those rules based upon the examination
of additional data. The patterns and rules might provide guidelines for decision making
or they might identify the issues upon which the decision maker should focus during the
choice process.
MODEL-BASED MANAGEMENTSYSTEMS
The DSS provides the decision maker with more than the models themselves. Through
the Model Base Management System (MBMS), the DSS provides easy access to models
and help in using those models. Clearly, the library of models is an important aspect of
this component. Such a library should provide decision makers access to a wide variety
of statistical, financial, and management science models as well as any other models that
could be of importance to the particular problems to be encountered.
Easy Access to Models
The library of models is provided so as to allow decision makers
easy
access to the models.
Easy access to the models means that users need not know the specifics of how the model
runs or the specific format rules for commanding the model. For example, consider the
screen from the SAS Data Miner module, shown in Figure 4.14. In Figure 4.14, we can see
that users can easily select a model simply by clicking on a tab shown at the top. In Figure
4.15,
which shows an application of IBM's Cognos, we see how the user can manipulate
the tools once they are chosen with simple keystrokes or mouse movements.
The MBMS should facilitate easy entry of information to the model. Unlike conven-
tional modeling software, which often requires that information be entered in a specific
order and a specific format, DSS should allow flexible input of the data. The role of the
MBMS is to translate the user-friendly form of the data into the appropriate format for a
Sentiment analysis is the effort to translate human emotion into data that can be used by decision
makers lo understand their
clients.
It
is,
in essence, the data mining of blogs and social networks to
examine and summarize reviews, ratings, recommendations* and other forms of personal opinion.
The tools attempt to categorize statements that are straightforward, such as"[ love this product" or
"I hate this movie," as well as those using sarcasm, irony, and idioms. Filtering through hundreds
of thousands of websites, these algorithms identify trends in opinions and some even identify
influential opinion leaders. Such tools could help companies pinpoint the effect of specific issues
on customer perceptions, helping them respond with appropriate marketing and public relations
strategies. For example, when Lhere was sudden negative blog sentiment against the Yankees, they
turned to sentiment analysis to identify the issue. The sentiment analysis identified a problem
associated with
a
rain-de laved Yankees-Red Sox game. Stadium officials mistakenly told hundreds
of fans that the game had been canceled, but their electronic ticket vendor denied fans' requests
for refunds on the grounds that the game had actually been played. Once the issue had been
identified, the company offered discounts and credits to the affected fans
and
reevaJuated its bad
weather policy.






160
MODEL COMPONENT
Figure 4.14. Simple model selection. Copyright © 2009, SAS Institute Inc. All rights reserved. Image reproduced
with permission of SAS Institute Inc., Cary NC, USA. (Source: http://www.sas.com/presscenter/screenshots.html.)
particular model. For example, even if a model requires the data be input in a rigid line and
column framework, such as shown below,
1.22 15 3
2.31 21 6
3.11 11 9
the user can input them (if they are not already in a database) flexibly in a format that might
be more comfortable, such as 1.22, 2.31, 3.11, 15, 21, 11, 3, 6, 9. The MBMS will put the
data in the format appropriate for the particular model(s) being used.
Similarly, users of the system need not be aware of the specific syntax required to
execute a particular model. The MBMS should generate the necessary commands to tell
the machine where the model is located and what commands are necessary to cause the
model to execute. For example, the user should not need to remember (or even know) the
requirements for naming or formatting the data to utilize them in a model. Rather than
the user needing to remember the code, such as that shown in Code 4.1, the user would
simply "click" on the icon for accounts data. Clearly, someone would need to program the
system to associate a particular icon with a given place in the database. More important






MODEL-BASED MANAGEMENTSYSTEMS
161
Figure 4.15.
Simple manipulation of a model. Screen shot from the 4.02 mark of the Cognos
video,
"Forecasting in Turbulent Times":http://www-01.ibm.com/software/data/cognos/solutions/
software-reporting-analysis.html; Image is reproduced courtesy of International Business Ma-
chine Corporation http://download.boulder.ibm.com/ibmdl/pub/software/data/sw-library/cognos/
demos/od.forecasting/rollingforecasts.html
.
from the perspective of the MBMS, though, is the fact that the data have been identified in
the appropriate format as input to a particular package (in this case, SAS).
Further, it is important that the program be notified that there is something "unusual"
about the data, such as the record length. Not only might users be unaware of the appro-
priate syntax through which to share this information, they might not even know that the
information needs to be provided. Similarly, users should not need to remember the control
sequences for testing hypotheses (Code 4.2); they could simply type
is there a difference
in absenteeism in the different groups?
Of course, in order to provide this easy access to models, the designer must make
certain assumptions about how the decision makers want their analyses conducted. In this
case,
the designer made assumptions about the specific test of the differences of means
among the groups by specifying the model, the test, the procedure, and the format of
output. On the one hand, this makes analysis easier for the decision makers because they
can access the model immediately without needing to specify assumptions, look up syntax,
Code 4.1 Sample Code to Input Data from a Modeling Package
CMS FILEDEF ACCOUNTS DISK ACCOUNT DATA AI (LRECL 135);
DATA SAMPLE;
INFILE ACCOUNTS;
INPUT DEPARTMENT $ 1-7 EMPLOYEE $ 9-2
5
NUMBER 27-32
ABSENT_FUL
L 34-36 ABSENT_HALF 38-42 REASON 80-133;
TOT_ABSENT = ABSENT_FULL+ABSENT_HALF
;






162
MODEL COMPONENT
or write code. On the other hand, it constrains those decision makers who need different
assumptions for their particular test. This presents somewhat of a dilemma for the designer
of the system in knowing how to make the trade-off between flexibility and control.
Regrettably, there is not a standard answer to this question, and only knowledge of the
decision makers, their preferences, their agreement on their preferences, and the likelihood
of their changing preferences will define how much flexibility is needed in the model
features. However, a designer can compromise. If, for example, most decision makers want
the features set in a particular way but not all accept this option, the features could be set
with a default setting and easy access to change the settings. Upon the selection of the test,
a window such as that shown in Figure 4.16 could appear. As the users click a mouse (or
press enter) on any one of those, they would see another window that allows them to change
the options.
There are variations on this approach. If, for example, the differences in features
is person specific, the designer could build intelligence into the system with a rule that
specifies that, if the user is PERSON X, the Gabriel test rather than the Duncan test should
be used. In this way, PERSON X always has the preferred test as the default and all others
have their preferred test as the default. Or, the designer could provide a check box that
Figure 4.16. Model option selection.
Code 4,2 Sample Code to Process Data from a Modeling Package
PROC ANOVA;
CLASS 0FFICE1 OFFICE2 OFFICE3;
MEANS 0FFICE1 OFFICE2 OFFICE3 OFFICEl*OFFICE3/DUNCA
N LINES;
MODEL Y = OFFICEl | OFFICEl*OFFICE3 | 0FFICE3/INT INTERCEPT;
TEST H = TOT^ABSEN
T TOTJ^SENT*SENQRITY JOB
TITLE 'ABSENTEEISM BY OFFICE, SENIORITY, JOB';
CLASSES
MEANS
MEANS TEST
MODEL
INTERCEPT
HYPOTHESIS TESTS
Hr
E:
defined as A B C
specified as A B C andA'C
Duncan
defined as main effects
tested as an effect
will be printed
specifies numerator effects
specifies denominator effects






MODEL-BASED MANAGEMENTSYSTEMS
would allow users to change defaults before running the test if they desire. While it is
tempting to force the user to acknowledge and accept each option individually, it is not
recommended. Such a sequence will increase the average amount of time it takes for a
user to run a model. Unless many users often change the options, this is an unnecessary
waste of time. In addition, many users will quickly tire of these repeated entries, learn to
ignore them (by pressing accept for each option), and become frustrated with the system.
Furthermore, they will not be any more likely to actually read the entries.
Understandability of Results
In addition, the DSS should provide the results back to the user in an understandable form.
Most models provide information to the user employing at least some cryptic form that is
not comprehensible for people who do not use the package frequently. For example, the
results from a regression could be presented using a standard output format of a commonly
used modeling package, such as that shown in Figure 4.17. Regular users of this modeling
package can find most of the information that they need to evaluate the model and begin
forecasting with it. However, even a person who is familiar with statistics but unfamiliar
with the output of this package or other statistical packages might not be able to interpret
the meaningfulness of the results. Certainly a decision maker not familiar with either
statistics
or
the modeling package would be unlikely to be able to answer even the simple
question of how many items one would expect to sell if the price were $1.24 and the
advertising expenditures were $15/month. Consider, instead, a screen such as that shown in
Figure 4.18.
In Figure
4.18,
the results are labeled clearly and all the relevant information
is
provided
to the user in a conclusion format. The user does not need to remember too much about
the technique "regression" because the screen explains the types of issues that should be
of interest. Furthermore, it encourages the user to experiment with the model (by entering
data) so as to become more comfortable with it and the
results.
Since one of the fundamental
assumptions in DSS design is that the supported decisions are "fuzzy" and infrequently
encountered, it is important not to assume that the user can remember the nuances of the
output of each model that might be accessed.
Note that we are not simply talking about the
appearance
of
the
results.
In Figure 4.18,
we are literally helping the user to understand the
meaning
of
the
output by removing some
of the jargon implicit in the computer printout and rephrasing in terms the decision maker
can understand. For example, consider the boxed information on the left. The purpose
of the box is to highlight the meaning of the slope coefficient associated with each of
the variables as well as their associated interval estimates. In contrast, Figure 4.17 lists
the slope in the column "parameter estimate" next to the respective variable name. The
appropriate standard error appears in the following column. To use the information from
the the modeling package output, the decision maker needs to know what each of these
terms means and that a slope can have a physical interpretation. Furthermore, the decision
maker needs to know that all point estimates have intervals associated with them and that we
determine the interval by multiplying the standard error by the critical value of
t
associated
with 48 degrees of freedom, which is found in a standard
t
table but not in Figure 4.17!
This is a lot to expect from the decision maker, especially given that each model has its
own unique notation and set of
issues.
The box in Figure 4.18 does not require the decision
maker to know all the intermediary steps or to compute anything. In short, Figure 4.17
provides results from the model. Figure 4.18 provides
support
for a decision.
Clearly, different individuals will require different levels of support. Figure 4.18 pro-
vides only the minimal quantitative information. However, it can be tied to other output






164
MODEL COMPONENT
Figure 4.17. Traditional results format.
screens that could provide additional support if the decision maker selects it. For exam-
ple,
in Figure 4.21, the instructions note that the user can obtain additional information
about a specific topic by clicking the mouse on that statement. In this screen, the statement
"both variables are useful" is highlighted. If the decision maker clicked on that space, the
system would display Figure 4.19, which provides additional information, including the
mathematics and assumptions behind the statement.
The previous example provides information to the decision makers only if they select
it. However, sometimes you want to make sure that the decision maker sees additional
help screens because it is crucial. In this case, the system can "force" a particular area of
the screen to be highlighted, create a "pop-up" notice about a problem, or emit a sound
to catch the decision maker's attention. Suppose, for example, the variable "price" in the
model described in Figure 4.20 were
not
statistically significant. It is possible to provide
the information in a box as shown in Figure 4.20.
This box provides information about the validity of the model. However, it is passive
and does not highlight the problem or tell the decision maker the implications of the
problem. Instead, consider Figure 4.21. In this screen, we are highlighting some of the
information so that it is not missed by the decision maker. Not only does this additional
screen call attention to the easily missed note about the variable being not statistically
significant, the "CAUTION" screen tells the decision maker the implications of not taking






MODEL-BASED MANAGEMENT SYSTEMS
165
Figure 4.18. Results with decision support.
action on this problem. In this way, the DSS is helping the clients clarify their assumptions
about the implications of the results. So, in fact, the DSS is helping the decision maker to
use the information correctly.
The way we accomplish this task depends on what kind of DSS generator and modeling
package we are using. In an abstract sense, there must be code that causes the computer to
scan the results of the model and creates the base screen with the results. In this case, the
modeling package must return the results of the
F
statistic, the
t
statistics, the probabilities
associated with those
t
statistics, and the mean squared error. Further, there must be some
"intelligent" link that fires to interpret the results and to place those results in the appropriate
window. Finally, there must be another intelligent link that fires when one of the variables
is not significant to cause the "CAUTION" screen to appear.
Clearly, creating this kind of help in a traditional language is difficult. The fourth-
generation languages and object-oriented languages available today allow the designer
much more flexibility. First, such languages allow the user to create "pop-up" windows
that are linked to particular results or variables. In this case, each of the four items noted
in the results window might actually be a different window that is linked to code checking
the appropriate result. The border might actually be a hyperregion that serves no purpose
but an aesthetic one. Furthermore, the "CAUTION" screen might be linked to an indicator
of nonsignificance of a variable. An alternative "CAUTION" screen might be linked to a
condition where two or more of the variables are not significant.






166
MODEL COMPONENT
Figure 4.19. Detailed model support.
Integrating Models
Another task of the MBMS is to help integrate one model with another. For example,
suppose the user needs to make choices about inventory policy and selects an economic
order quantity (EOQ) model, as shown in Computation
4.1.
To
use this formula
to
determine
the optimal order quantity, we need information about expected product demand, the costs
associated with an order, and the typical holding costs (with consistent monetary and time
units).
If the decision makers can input the data or read the data directly, there is no problem.
Typically, however, this is not the case. Generally, the order costs need to be computed by
combining the costs of personnel, supplies (such as forms), and services (such as phone
resources) needed to execute an order. In addition, since holding costs can vary over time,
we need to average holding costs to obtain a current estimate. Finally, unless demand is






MODEL-BASED MANAGEMENTSYSTEMS
167
Figure 4.20. Passive warning of model problems.
well specified, it needs to be forecasted based on historical data. Hence, upon selection of
the EOQ model, the MBMS needs to complete several tasks:
1.
Search the database for a single value for the order costs.
2.
If
no
specific order cost information is available, invoke the model to compute order
costs by summing personnel costs, supply costs, service costs, and the order cost
charged by the vendor.
3.
Feed the computed order costs to the EOQ model.
4.
Obtain data about holding costs.
5.
If historical data are available, estimate holding costs.
6. If no historical data are available, invoke the model to determine holding costs.
7.
Feed the computed holding cost value to the EOQ model.
8. Invoke the model to forecast demand for the time period(s) served by the order.
9. Feed forecasted demand to the EOQ model.
10.
Compute the economic order quantity.
The user not only should not need to intervene in this process but also should not
need to know the process is occurring. However, since the meaningfulness of the EOQ is
dependent upon the quality of the forecasts and estimates, the user should be provided the






168
MODEL COMPONENT
Figure 4.21. Active warning of model problems.
forecasts and information about the quality of those forecasts. This might be accomplished
as in Figure 4.22.
Sensitivity of a Decision
One of the tasks of the model base management system in a DSS is to help the decision
maker understand the implications of using a model. This is not always easy because
decision makers may not be inclined to ask questions, particularly if they do not know what
questions need to be asked. Consider the following examples.
Example 4.1. Peara's Personalized Widgets uses an assembly line to build desired con-
figurations. One of the employees on the line has suggested a change in procedure that
Andrew Peara thinks might improve the efficiency of the operations. Andrew Peara wants
to determine if his intuition is correct and if the change would be worth implementing. To
investigate this using historical data, he determines that the mean length of time to perform
a certain task or a group of tasks on an assembly line is 15.5 minutes, with a standard de-
viation of
3
minutes. Because he understands the importance of collecting data, he selects
16 employees and teaches them the new procedure. After a training period, he finds these
employees, on average, take 13.5 minutes to perform the task with the new procedure.
The question Andrew needs to answer is whether these results provide sufficient evidence






MODEL-BASED MANAGEMENTSYSTEMS
169
Figure 4.22. Integration of models.
to indicate that the new procedure is really faster and thus should be implemented. This
statistical analysis for this problem is shown in Computation 4.2.
Computation 4.2. Sample f-Test
In introductory statistics, you learned that this type of problem is a one-tailed test of the mean.
From a statistical point of
view,
the question is
Ho:
H
A
:
μ=
15.5
μ < 15.5
Where μ is the true mean task time. To test this, given the sample size of 9 and the estimated
standard deviation, one uses a
t
test: Reject
HQ
if computed
t
is
less
than the critical
t
value,
ti£ =
1.8331, or
t
=
μ
tf^fn
<
1,8331
In this problem,
t
=
■
ß
13.5-15.5
= -2
Since the calculated value for /is less than the critical value of
t
(in standard r tables found as
—
1,8331
)>
one can reject the null hypothesis,






MODEL COMPONENT
Based on the analysis, Andrew Peara knows that there is reason to believe the new
procedure will reduce the amount of time it takes to perform the task. However, it is unlikely
that this is the only information the decision maker will want to know in order to make the
decision. It is obviously necessary to determine whether the value of
the
additional widgets
that could be produced (because it takes less time to perform each one) offsets the cost
associated with the training. We could then estimate that instead of producing 3.87 widgets
per hour (one every 15.5 minutes), the average person will be able to produce 4.44 widgets
per hour. Said differently, this is an increase of 4.59 widgets per shift, or 22.98 widgets per
week for an average worker. With this information and some information about the revenue
per computer and the cost of training, the decision maker can easily decide whether the
additional 23 widgets per week per worker will increase revenue sufficiently to justify the
costs of training.
However, this analysis is built upon some assumptions that may not be clear to the
decision maker. One of the characteristics of good decision support is that it helps the deci-
sion maker understand these assumptions and evaluate whether or not they are reasonable.
Before discussing how to display the information, we need to know the assumptions.
1.
A major assumption underlying this analysis is that these 16 individuals really do
represent the employees who will perform this
task.
While the description of the
problem indicated that the 16 were "randomly selected," it is important to be sure
they are representative. In real-world cases, "randomly selected" might mean the
16 people who volunteered, the 16 best workers, the 16 biggest problems for the
supervisor, or the 16 people who happened to make it to work on a very snowy
day. Since you are not provided with information regarding how the sample was
selected, it is important to test whether these employees really were representative
by comparing their task times prior to the introduction of the new procedures to
their times afterward (such as through a paired
t
test).
Consider the three possibilities and how they could affect the decision. If their
average pretraining assembly time were not statistically different from that of the
entire group, then the original conclusion appears valid. If instead their average
pretraining task time were statistically larger than that of the group, the results
are potentially more impressive. This fact should be brought to the attention of
the decision maker as even more evidence that the training is good. However, if
their average pretraining assembly time already were statistically lower than 15.5
(especially if it is statistically lower than 13.5), Andrew Peara would need to know
the training might not be as effective as the test first indicated.
2.
A second assumption is that the variance associated with task completion will
not be increased.
The original description of the case indicated that the standard
deviation is 3 minutes. Since one of the major causes of bottlenecks on assembly
lines is increases in variation of assembly time, it is necessary to determine whether
the posttraining standard deviation is still 3 minutes. Problems in balancing the
line and/or quality control will almost certainly occur with an increase in the
variance.
3.
One of
the
basic assumptions is that there is demand for the extra capacity.
The
benefits of achieving this new efficiency can only be realized if either there is
demand for additional items or the workers can be used profitably in some other
task. If
not,
regardless of
the
results of the test, incurring the cost of the new training
is not worthwhile.






MODEL-BASED MANAGEMENTSYSTEMS
As with most aspects of decision support, there is no universally correct way to provide
this information to the decision maker. The basic options are (a) check the assumptions
automatically and note the results on the screen in a pop-up box; (b) check the assumptions
automatically and only note the violations of the assumptions on the screen; (c) note the
assumptions on the screen and allow users to decide whether or not they need to be checked
(either individually or as a group); or (d) ignore the assumptions and assume the users
know enough to check them without system help. Clearly each option has advantages and
disadvantages. If we provide total information (the results of the tests on the screen), then
the user is informed about the reasonableness of the use of the statistic. However, users
may find this information clutters the screen, especially if many assumptions are evaluated
for a given test. In addition, users may not take the time to scan the information box and
hence may not notice the violations. Similarly, if we simply give the users the option of
checking assumptions, they may not take the time because they do not know the value of
the additional information. However, if the users are quite knowledgeable about their data,
this option saves processing time and hence provides a faster response to the user. By not
warning the users of the potential problems, we fail to provide decision
support.
The remaining option, check the assumptions and list only those that are not validated
by the check, provides the support necessary to help users apply the techniques better. In
addition, since only problems are noted on the screen, the results do not become tedious and
users know they should pay attention to them. Of course, testing the assumptions can use
more processing time and hence slow response time. If this is perceived to be a problem,
we can always allow the user to set options to ignore the testing of
one
or more assumptions
prior to running the test. We even can build these preferences into a profile for each user so
they do not need to be set each time a model is invoked.
In addition to testing assumptions to verify that a model is being appropriately used,
the decision maker might simply want to develop a better intuition for the problem. The
MBMS should help users to investigate more facets of a problem easily. Typically, such
additional analyses are menu options, not automatic procedures.
Consider the types of additional analyses that might be undertaken in the problem of
the mean task times just considered. Clearly, additional analyses are more crucial if the
results of the analysis suggest there is no difference in the two means. Such intuition can
be facilitated by the system giving information about the sensitivity of the results to the
various conditions of the problem. For example, it might be quite reasonable to provide
some information about what mean time would be necessary to produce a statistically
significant result. This can be determined by using the same equation but solving for the
sample mean necessary to achieve the critical value of
t
(from a statistical table), as shown
in Computation 4.3. So, as long as the new procedure takes, on average, less than 13.67
minutes, it will produce a statistically significant improvement. Alternatively, we might
want to know how large a sample would have been necessary to obtain significance with
the result of an average time of 13.5. Again, it is simply an issue of considering the base
formula in a slightly different manner as shown in Computation
4.4.
In this case, the results






172
MODEL COMPONENT
suggest that it was only necessary to have three subjects with the data that are available.
If the test had not been significant and Andrew Peara would want to rerun the test with a
different number of subjects, this equation would tell him how many subjects to select.
Example 4.2.
Consider a second example where a decision maker selects regression to
help solve a problem. In this case, a manufacturer wants to know the relationship between
the age of machinery and the annual maintenance costs. A sample of 50 machines is taken
and the following costs are obtained:
Age
(months)
Ϊ
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
1
20
Maintenance
Costs
81
35
114
36
91
134
45
130
170
141
188
145
220
119
134
196
154
207
188
226
Age
(months)
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
Maintenance
Costs
59
52
59
57
67
73
66
77
68
73
81
76
84
79
82
477
456
431
447
505
Age
(months)
41
42
43
44
45
46
47
48
49
50
Maintenance
Costs
543
457
491
588
596
602
580
654
559
678
If we constructed a screen for the results of this regression that paralleled that in Figure
4.21,
it would appear as that shown in Figure 4.23. It appears from the information provided in
Figure 4.23 that the model is good and should be used. However, this is not true. Although
the relevant statistical measures of the model have been checked and are significant, they
do not convey the complete story about the implications of using this model. Consider the
graph of the maintenance data shown in Figure 4.24. With a quick examination of the data,
it becomes obvious that there is some phenomenon occurring in the middle of the data.
This change in process is undoubtedly affecting the equation. More importantly, from a
prediction point of view, of course, is the fact that the equation is not particularly good






MODEL-BASED MANAGEMENT SYSTEMS
173
Figure 4.23. Modeling results with some interpretative support.
Figure 4.24. Plot of maintenance data.






174
MODEL COMPONENT
Figure 4.25. Model results with better interpretative support.
at predicting costs for those 10 machines. This suggests that the age of the machinery is
not sufficient to determine maintenance costs and that some other phenomena need to be
considered. From the user's perspective, the graph suggests that while age might be a good
indicator in general, it is necessary to understand the maintenance issue better.
It is difficult, even with today's technology, to have the computer scan the graph and
alert the decision maker to problems in the data. Since the graph conveys information not
communicated by the statistics, it is useful to provide a way for decision makers to get
to the graph easily. If the decision makers can be relied on to look at the information,
simply providing the ability to view the graph through a click of a button is sufficient. An
alternative is to have the graph be part of the screen, as shown in Figure 4.25.
Model Management Support Tools
The kinds of issues associated with model-generated questions like those in the two exam-
ples will, of course, depend upon what model is being used. For example, if the decision
maker is using linear programming to determine a mix of products to produce with a lim-
ited set of inputs, then sensitivity analyses will include questions such as: (a) what if the
company has more of a particular input than specified; (b) what if the company has less of
a particular input than specified; (c) what is the impact on production policies if the price






MODEL-BASED MANAGEMENT SYSTEMS
175
Figure 4.26. Passive prompting for further analysis.
of an input changes; (d) what is the impact on production policies if the selling price is
changed; and (e) what is the impact if we change the relative input needs of the possible
products? Alternatively, if we are using a financial analysis, the questions might be, "How
is present value affected by discount rate, tax rates, or depreciation.?"
Further analyses also might be prompted by a particular result of an analysis. For
example, suppose that the DSS has been created
to
support marketing research for
a
clothing
manufacturer. Suppose further that someone found a result that the demand for the high-end
trousers was declining in some states but increasing in other states. This might prompt the
decision maker to ask questions, such as what do the states where sales are increasing have
in common and what do the states where sales are decreasing have in common. In particular,
the decision maker might be interested in the demographic distribution of the states, the
distribution of competitors in the states, and the similarities in income, population, industry,
or metropolitan areas in the states. Hence, for the system to be effective, the decision maker
should be able to query it about each of these facts. Suppose that in these queries the
decision maker finds the average age of white collar workers is higher in the states where
the trousers are selling well than the states where the trousers are selling poorly. This
provides the decision maker with some information. Perhaps the company officials already
know that their product appeals to more mature clientele. Then, the results probably will
not be investigated. However, if decision makers perceive the product appeals more to
younger clientele, then this information would suggest a need for further modeling to test
the underlying assumptions of their market research efforts.






176
MODEL COMPONENT
Figure 4.27. Active prompting for further analyses.
Perhaps, upon receiving the information regarding declining sales, the decision maker
who is new has no theories about what could be happening. A good DSS should be able
to help those decision makers work through the analyses. For example, it should be able
to prompt the decision maker to consider issues such as demographic changes in the area,
employment trends, costs of living, and other factors specific to that particular product.
Such help might come in terms of a simple "why" key available on the screen, as shown
in Figure 4.27. Or, it might allow appropriate information boxes to appear, such as shown
in Figure 4.28. Alternatively, the decision maker might want to know how the trends are
expected to change over the next five years. Another screen might provide information
about expected trends.
The important aspect of this kind of support is to provide enough of the appropriate
information for the decision maker to understand the phenomenon of interest. The "WHY?"
key might provide information about automatic analyses among predefined options and
display them on the screen. In this way, the decision maker could click a mouse on a
particular statement and identify the appropriate analyses that generated it. The result of
this action might be the display of all related analyses or it might simply be the display
of all significant related analyses. Although each option is appropriate in some cases, a
general rule for selecting between these options is: The higher in management or the less
statistically trained the person, the less nonsignificant analytical results the DSS should
show.






CAR EXAMPLE
177
Figure 4.28. Assistance for defining criteria.
Or the "HELP!" key might provide information about the kinds of analyses that might
be accomplished to further investigate the topic. This differs from the "WHY?" option in
that it allows the decision maker to explore the relationships through whatever analyses are
deemed appropriate. With the "WHY?" option, the user is provided "canned" analyses to
consider. Alternatively, with this option, the system recommends analyses but allows the
user to select either one of the recommended or user-defined analyses. Such an option can
allow an unknowledgeable decision maker to learn more about the decision environment.
It can also allow the very knowledgeable decision maker to pursue some subtle clue that is
suggested by some earlier result.
CAR EXAMPLE
A careful consideration of models for the DSS could result in a system that allows users to
make truly informed decisions. Models should provide support for all phases of decision
making, from the initial generation of alternatives to the final questions of how to finance.
In addition, the model management component should include assistance in the appropriate
use of models and intelligence that reviews model use for errors. Finally, where possible,
the model management system should implement heuristics to automate choices where
decision makers cannot or do not implement choices.
Brainstorming and Alternative Generation
One important model management operation is to help users generate alternatives. At the
simplest level, alternative generation could include searching for options that meet some






MODEL COMPONENT
criterion specified by the user. Some users will want a car that looks "cool" and goes fast.
Others will want a car that will facilitate their car-pooling activities or that will be good for
trips.
Still others will want to consider fuel efficiency or safety in their analysis. Others will
just want a car they can afford. The search process is straightforward and was illustrated in
the previous chapter.
More likely scenarios, however, are that the user is not sure about the criteria he or she
wants to employ or that the user has a general idea of the criteria but does not understand
the specific factors to employ. The DSS should allow users to select any criterion or set of
criteria. However, if we put all possible criteria on a screen, users will find the interface both
difficult to read and overwhelming to use. If we put only a subset of the possible criteria
for consideration, though, we are making choices about the criteria that the decision maker
should
use—clearly an inappropriate function for a designer of a DSS. Even if we list
all possible criteria but use multiple screens to display them, we are suggesting a relative
importance of the criteria by the order in which they are listed.
Hence, the goal is to summarize and guide while still allowing a great deal of flex-
ibility. One possibility is to categorize criteria and ask users first to specify the
category
of criteria that they want to emphasize. For example, one could provide a menu choice
that includes the categories, such as comfort, convenience, financial, mechanical, safety.
Using this method, we could ask users to declare their criteria groups under the option
"criteria" as highlighted in Figure 4.28. If a user selected performance criteria (as is
highlighted), he or she would next select from factors that might be considered perfor-
mance criteria. This list might include items such as acceleration rates, horsepower, or
engine size since these items are clearly linked to performance. Others, however, might
consider factors such as fuel efficiency to be a performance characteristic, and so they
would be listed as well. At this screen, decision makers should be able to elect several
factors in a category. In this way, decision makers can continue to refine their choice
processes.
It is important to help users understand the implications of choices they select. One
part of such help is ensuring that the users comprehend the meaning of the terms used in
the questions. For example, suppose the user selected safety criteria from the screen shown
in Figure 4.28. The next screen to appear would be Figure 4.29. Notice in this figure there
is an icon for questions next to
each
criterion the users are asked to rate. So, if the user
did not know of the NHTSA or any of its ranking procedures, he or she could query the
icon next to NHTSA, and the system would respond with a pop-up box such as that shown
in Figure 4.30. This box would explain the NHTSA, document the rankings they perform,
and discuss the reliability and meaningfulness of its tests.
Another part of the model management function is to provide users with intelligent
help as they proceed through the system. For example, suppose a user selected
none
of the
factors listed in Figure 4.19. Since the system would be monitoring these selections, this
inaction would trigger the system to fire a demon that warns the user of inconsistency in
his or her choice of safety as an important criterion without selecting any individual criteria
against which the criteria would be evaluated. The kind of result one might get is shown in
Figure 4.31.
Rules such as these could be used in an evaluative manner as well. In this way, if
users select criteria that are likely to cause them problems, intelligent agents can give them
warning. For example, young, unmarried males tend to have very high insurance rates. So,
if such a person selected acceleration rate and engine size as the two most important criteria
(under the category of performance), then the system should respond with a warning about






CAR EXAMPLE
179
Figure 4.29. Finer detailed definition of criteria.
the cost of such a decision. This warning would be generated because the following rule
would be executed:
IF gender IS male AND age < 27 AND marital status IS single AND Performance
Criterion IS acceleration rate AND Performance Criterion IS engine size
THEN ASK warning display
This would result in a window such as that shown in Figure 4.32 to be displayed.
After the initial evaluations are completed, we might create a scratch sheet onto which
users could keep track of automobiles under consideration. A sample of
a
screen of this type
is shown in Figure
4.33;
this figure illustrates an actual screen from the commercial product
Axon, with a screen also showing creativity techniques. The goal is to have a scratch pad
onto which users can keep notes and the system can keep statistics.
Flexibility Concerns
Three possible problems are suggested with this plan. First, the user who already knows the
models of automobile he or she wants to consider will find this option difficult. Clearly it
is inappropriate to have these users go through the process of selecting general criteria and
specific factors and consider multiple automobiles so as to screen them down to
a
conclusion






180
MODEL COMPONENT
Figure 4.30. Content-dependent assistance of criteria selection.
upon which they have already arrived. Since they know the automobile or automobiles they
want to consider, the process should be straightforward. These users can use the "select"
option in the main menu that allows them to choose one or more automobiles directly and
proceed in the analysis from there.
A second problem is the user who wants to select a mixed strategy. This user wants
some characteristics specified under multiple categories. For example, the user might want
an automobile that has a high fuel efficiency as well as a good safety record. These users
also can be accommodated if the system allows them to move into other criteria categories
from the secondary screens. So, when the user has selected issues of importance under the
safety criterion, for example, he or she can then select an option of "identify other criteria"
and be given the list of criteria not yet selected, including comfort, convenience, financial,
mechanical, and performance, as shown in Figure 4.34.
The third problem
is
the user who has absolutely no idea of how to select an automobile.
In this case, the model management system should help users brainstorm criteria with
intelligent agents. Specifically, the system should invoke an expert system that focuses
on lifestyle questions and generates a set of criteria based upon the user's answers. The
system would ask users questions and process the answers based upon rules developed by
designers. For example, a rule such as
IF monthly disposable income < 200 THEN Criteria OF Preferences IS Financial






CAR EXAMPLE
181
Figure 4.31. Support for criteria definition.
Figure 4.32. Intelligent support in a DSS.






Figure 4.33. Brainstorming support tools.
(Source:
http://axon-research.com/axon/t_creative.gif. Designed by Brian Maskell, bmaskell@maskell.com,
http://www.maskell.com/lean_accounting/subpages/people/brian_maskell.html
.) Image reprinted here courtesy of Brian Maskell and Axon Research.






CAR EXAMPLE
183
Figure 4.34. Support for Multi criterion choices.
would tell the system to select financial criteria as paramount for those users who would
have difficulty making car
payments,
especially when coupled with maintenance, insurance,
and upkeep costs. However, another rule,
IF monthly disposable income > 1200 AND number of children 3 AND primary usage
IS car pooling
THEN Criteria OF Preferences IS Convenience
would tell the system to consider convenience criteria instead. While there is nothing
prohibiting these users from considering cost as a factor, the system would indicate that it
is not the primary criterion to be considered. In addition, the system should recommend
criteria that should not be applied to the selection of automobiles.
Evaluating Alternatives
As decision makers consider various automobiles, they compare the benefits and costs
associated with owning each of them. How they compare them depends upon the criteria
selected. For example, some decision makers might select the automobile that has the
greatest number of desirable features available at the lowest cost. Others may rely heavily
upon the performance statistics and feel of the drive. Still others may select the automobile
that comes most highly recommended by a trusted source.






184
MODEL COMPONENT
Figure 4.35. Specifying criteria.
Part of the modeling function of an automobile DSS is helping the decision maker to
compare those functions he or she thinks are important. As with the original definition of
the criteria, it is important to view these a limited number at a time. For example, consider
the screen taken from the commercial package
Auto Answers,
shown in Figure 4.35. A
very limited number of items are shown in this screen, all under the category "general."
As you can see in Figure 4.36, a dropdown menu allows users to select information from
a variety of categories. Each category gives information on a limited number of features
so as not to overwhelm the user. Of course, an improvement on this approach would
be to list the information for multiple alternatives in charts such as these. In that way,
users could
compare
the automobiles on the criteria of importance and see how they
relate. A system might, in addition, provide a relative score for each automobiles in each
category or a highlighting of that automobile that seems to provide better values on the
factors, so the user can easily see if there is a dominant alternative among the cars under
consideration.
Users might
also
want
the
opinion of trusted sources in the evaluation. Publications such
as
Consumer Reports, Kiplinger's Reports, Car and Driver,
or
Edmund's Guides
conduct
tests and rate automobiles in various areas. Tables such as that shown in Figure 4.37 could
be incorporated in the system. Users might want to couple this with raw access to text files
with reports on automobiles. An example is shown in Figure 4.38 which illustrates part of
Edmund's Guide
available on the Internet.
Another task for which the DSS could be helpful is in the estimation of the real
costs associated with the automobile. Generally, novice users who have not owned a car
previously examine only the car payments in an estimation of the cost. Consider the screen






CAR EXAMPLE
185
Figure 4.36. Results from analysis.
in Figure 4.39. through which the user is asked about his or her driving tendencies. To
respond to this inquiry, the system must complete the following tasks:
• Search the database for the desired model of automobile
• Query the database for fuel efficiency for highway and city driving
• Use the approximate miles driven (provided by the user) to compute the amount of
gasoline needed.
• Multiply the cost of gasoline by the amount of gasoline needed
• Compute the average monthly maintenance cost by dividing the expected annual
costs by 12
• Add together the maintenance cost and the gasoline costs
Using Cold Fusion, Javascript, and the Web, this could be accomplished with a program
such as that in Code
4.1.
The result of these operations can be found in Figure 4.40. The
DSS would serve the user considering multiple automobiles by providing the information
in tabular form coupled with historical information, such as that shown in Figure 4.41.
Models could also help the user with some of the most confusing aspects of purchasing
an automobile: financing. For example, they could be built to evaluate car prices under a
variety of financing alternatives. Consider the model shown in Figure 4.42. This system
allows users to explore the impact of various time periods for loans and various interest
rates upon the payment schedule. The choice of both time periods and interest rates would
be left for the user to specify. Once these are selected, the loan payment schedule table






186
MODEL COMPONENT
Figure 4.37.
Consumer Reports
data could be accessed from a DSS.
(Source:
http://www.consumerreports.
org/cro/cars/compare.htm?add=true&x=17&y=5&product=subaru%2Fimpreza&product=toyota%2Fcorolla%2Fle-4
-
cyl&product=ford%2Ffocus&product=suzuki%2Fsx4%2Fsedan-le-4-cyl&product=honda%2Fcivic%2Fsedan-gx-4-cyl.)
(bottom right) would be populated. If the user requests advice by pressing the "recommend
values" button, the system would respond with information about current interest rates and
loan periods at local financing institutions. In addition, the DSS could provide historical
trends and forecasts of future values. In this way, users can evaluate the impact of different
interest rates for different term loans, special rebates, free add on's, low down payment, or
no down payment.
The DSS should also provide intelligent assistance for these experiments by guiding
the
user.
For
example,
it could recommend sound sensitivity procedures such as maintaining
some variables constant from one experiment to the next. Since altering too many variables






CAR EXAMPLE
187
Figure 4.38. Edmund's car review.
(Source:
http://www.edmunds.com/toyota/corolla/review
.
html.) Copyright © 2009 Edmunds.com, Inc. Imaged reproduced with permission.
results in confusing analyses, the system should warn when such comparisons are being
conducted. For example, the user should be warned about comparing a four-year loan at
7%
to a five-year loan at 7.75% with a different down payment.
The ability to take into account the time value of money may provide a key tool to
some users. Some user's decisions may weigh heavily upon the net present value (NPV)
of a purchase rather than on the financing specifics of a purchase. Given this need, users
should be able to compare NPV results under a variety of purchase options.
Note that in Figure 4.43, the left side of the screen provides information about the
cost of the automobile. The information is a function of the automobile selected and the
options selected for that make and model of automobile. Since these selections were made
by the user on previous screens, it is important for the system to carry the values through to
this screen
automatically;
the user should not need to reenter the values or even remember
what they were. If the user wants to change the options or review the reasons for the cost,
he or she could select the "review" button and return to those screens from which the
selections are made. Similarly, the system should bring the information about likely dealer
discount from the database automatically as well as the information about taxes and fees. If
the system facilitates the trade-in of used automobiles, that information should be brought
forward as well.
The system might also help the user compare the outright purchase with a lease
agreement. It could help the user evaluate the options for lease most appropriate for his or
her specific needs. The user may be faced with options such as low or no interest given a
particular down payment or cash back instead of the special interest rates.






Figure 4.39. Queries like these are designed to help the user better understand his or her
choices.
Figure 4.40. Decision support results.






CAR EXAMPLE
189
Figure 4.41. Historical information to facilitate support.
Running External Models
Often we use external programs to obtain all of the modeling support we need. There are
a variety of ways of implementing models depending upon the environment in which one
is operating. On the one hand, integration may be simply facilitating the user's access to
external modeling packages. For example, suppose decision makers needed access to the
package Excel to facilitate a variety of kinds of modeling, especially when the spreadsheet
has embedded macros. Using Javascript, designers could create a push-button that invoked
the following code:
<FORM>
<A HREF= "analysis.xls"><INPUT TYPE="BUTTON" VALUE="VIEW
ANALYSIS" ></A>
</FORM>
This code will cause a batch file that sets the appropriate environment settings that allows
Excel to run and to open a spreadsheet called "analysis." If the spreadsheet were invoked
with macros running, appropriate data could be accessed automatically, and
the
user could be
led through particular analyses using just
the
functions in the
macros.
Through those macros,
designers could build useful model management functions similar to those discussed in this
chapter. Of course, similar functionality could be included with other external modeling
packages.






190
MODEL COMPONENT
Figure 4.42. Support for users exploring assumptions.
DISCUSSION
The goal of the model management component of a DSS is to help decision makers
understand the phenomenon about which they are making a choice. This involves helping
them to generate alternatives, measure the worth of those alternatives, and make a choice
among those alternatives. In addition, the model management component should have tools
that help the decision maker use the models and evaluate the results effectively. Designers
need to include both passive and active assistance for the decision makers. Context-specific
help for using and interpreting models needs to be available for the user. In addition, the
system needs to monitor violations in the assumptions of models or irregularities of their
use and bring them to the attention of the user. Finally, all of this support should happen
in a manner that is easy for the decision maker to understand and not threatening from a
technical point of view.
SUGGESTED READINGS
Acquisiti, A. And O. Povel, "Predicting Social Security Numbers from Public Data,"
Proceedings of
the National Academy of Sciences of
the
United States of America,
Vol. 106, No. 27, July 2009,
pp.
10975-10980.
Aldrich, C.
The Complete Guide to Simulations and Serious Games: How the Most
Valuable
Content
Will
be Created in the Age Beyond Gutenberg to Google,
New York: Pfeiffer, 2009.






SUGGESTED READINGS
Auclair,
P.
F, S. J. Wourms, and J. J. Koger, "Ideas Take Flight,"
ORMS
Today,
Vol. 20, No. 4, August
1993,
pp. 24-29.
Baker, S.,
The Numerati,
New York: Mariner Books, 2009.
Baldwin, A. A., D. Baldwin, and T. K. Sen, "The Evolution and Problems of Model Manage-
ment Research,"
Omega: International Journal of Management Science,
Vol. 19, No. 6, 1991,
pp.
511-528.
Beemer, B., and D. G. Gregg, "Advisory Support in Decision Making,"
Handbook on Decision
Support
Systems,
Vol. I, Berlin: Springer-Verlag, 2008, pp. 511-528.
Betts,
M., "Efficiency Einsteins "
ComputerWorld,
Vol. 27, No. 12, March 22, 1993, pp. 63-64.
Bhargava, H. K., and S. O. Kimbrough, "Model Management: An Embedded Languages Approach,"
Decision Support Systems,
1993, pp. 277-299.
Bonczek, R. H., C. W. Holsapple, and A. B. Whinston, "The Evolving Roles of Models in Decision
Support Systems,"
Decision Sciences,
Vol.
11,
No. 2, 1980, pp.
616-631.
Brightman, H. J.,
Statistics in Plain English,
Cincinnati, OH: South-Western Publishing Company,
1986.
Brown, G. G., and R. E. Rosenthal, "Optimization Tradecraft: Hard-Won Insights from Real-World
Decision Support,"
Interfaces,
Vol. 38, No. 5, September-October, 2008, pp. 356-366.
Buchs,
M.
And
P.
Hättenschwiler, "Model Inspection
in the
Context of Distributed
DSS,"
International
Journal of Decision Support
Technology,
Vol. 1, No. 4, October-November, 2009, pp. 16-37.
Butters, S., "Jewish Hospital Healthcare Services Uses DSS,"
Journal of Systems Management,
Vol.
43,
June 1992, p. 30.
Climaco, J., J. Costa, L.C. Dias, P. Menlo, "Supporting Collaborative Multi-Criteria Evaluation: The
VIP Analysis Plug-In for Decision Deck,"
International Journal of Decision Support Technology,
Vol. 1, No. 4, October-November, 2009, pp. 1-15.
Coy, P., and R. D. Hof, "3-D Computing: From Medicine to War Games, It's a Whole New Dimen-
sion,"
Business
Week,
September 4, 1995, pp. 70-77.
Davenport, T. H., and J. G. Harris,
Competing on Analytics: The New Science of
Winning,
Boston:
Harvard Business School Press, 2007.
DeOrio, A. and V. Bertacco, "Human Computing for EDA"
Design Automation Conference (DAC),
San Francisco, CA, July 2009.
Dodds, P. S., and C. M. Danforth, "Measuring the Happiness of Large-Scale Written Expres-
sion: Songs, Blogs and Presidents,"
Journal of Happiness Studies,
July 17, 2009, available:
http://www.springerlink.com/content/757723154j4w726k/fulltext.html.
Fedorowicz, J., and G. B. Williams, "Representing Modeling Knowledge in an Intelligent Decision
Support System,"
Decision Support Systems,
Vol. 2, No. 1, pp. 3-14.
Ferris,
J, "How to Compete on
Analytics:
The Analytical Center of Excellence," A SAS Institute White
Paper, available: http://www.sas.com/apps/whitepaper/index.jsp?cid=6426
, nd, viewed January
29,
2009.
Froelich, J., and S. Ananyan, "Decision Support via Text Mining,"
in Handbook on Decision
Support Systems,
F. Burstein and C.W. Holsapple (Eds.), Vol I, Berlin: Springer-Verlag, 2008,
pp.
609-634.
Gelman, A., and D. Weakliem, "Of Beauty, Sex and Power,"
American
Scientist,
Vol. 97, No. 4,
July-August 2009, pp. 310-314.
Geoffrion, A. M., "Computer-based Modeling Environments,"
European Journal of Operational
Research,
Vol. 41, 1989, pp. 33-45.
Gillespie, T. W., J.A. Agnew, E. Mariano,, S. Mossier, N. Jones, M. Braughton, and J. Gonzalez
"Finding Osama bin
Laden:
An Application of Biogeographic Theories and Satellite
Imagery,"
MIT
International Review,
February 17, 2009, available: http://wefrmit.edu/mitir/2009/online/finding-
bin-laden.pdf.






MODEL COMPONENT
Giraud, C. And O. Povel, "Characterizing Data Mining Software,"
Intelligent Data Analysis,
Vol. 7,
No.
3,
2003,
pp. 181-182.
Hagtvedt, R.,
P.
Griffin, P, Keskinocak,
P.
and
R.
Roberts, "A Simulation Model to Compare Strategies
for the Reduction of Health-Care-Associated Infections,"
Interfaces,
Vol. 39, No. 3, May-June,
2009,
pp. 256-270.
Han, H., and M. Kamber,
Data Mining Concepts and
Techniques,
2nd ed., Burlington, MA: Morgan
Haufmann, 2005.
Hillier, F. S., and G. J. Lieberman,
Introduction to Operations Research,
4th ed., Oakland, CA:
Holden-Day, 1986.
Holsapple, C. W., and A. B. Winston,
Decision Support Systems: A Knowledge-Based
Approach,
St.
Paul, MN: West Publishing, 1996.
Hollywood, J., K. Strom, and M. Pope, "Can Data Mining Turn Up Terrorists?"
ORMS Today,
Vol. 36, No. 2, February 2009, pp. 20-27.
Huff,
D.,
How to Lie with Statistics,
New York: W.W Norton, 1954.
Huh, S. Y., "Modelbase Construction with Object-Oriented Constructs,"
Decision Sciences,
Vol. 24,
No.
2, Spring 1993, pp. 409-434.
King, D. L., "Intelligent Support Systems: Art, Augmentation and Agents," in R. H. Sprague, Jr. and
H. J. Watson (Eds.),
Decision Support Systems: Putting Theory into Practice
3rd ed., Englewood
Cliffs,
NJ: Prentice-Hall, 1993, pp. 137-160.
King, M. A., J.
F.
Elder, B. Gomolka, E. Schmidt, M. Summers, and K. Toop, "Evaluation of Fourteen
Desktop Data Mining Tools,"
IEEE International Conference on Systems, Man, and Cybernetics,
San Diego, CA, October 12-14, 1998, pp. 2927-2932.
Konsynski, B., and D. Dolk, "Knowledge Abstractions in Model Management," in P. Gray (Ed.),
Decision Support and Executive Information Systems,
Englewood Cliffs, NJ: Prentice-Hall, 1994,
pp.
420^37.
Larose, D. T.,
Discovering Knowledge in Data: An Introduction to Data Mining,
New York: 2005.
Le Blanc, L. A., and M. T. Jelassi, "DSS Software Selection: A Multiple Criteria Decision Method-
ology," in R. H. Sprague, Jr. and H. J. Watson (Eds.),
Decision Support Systems: Putting Theory
into Practice
3rd ed., Englewood Cliffs, NJ: Prentice-Hall, 1993, pp. 200-114.
Leinweber, D. J.,
Nerds on
Wall
Street,
New York: Wiley, 2009.
Lichtman, A.,
The Keys to the White House: A Surefire Guide to Predicting the Next President,
Lanham, MD: Rowman & Littlefield, 2008.
Lieber, R., "American Express Kept a (Very) Watchful Eye on Charges,"
New York Times,
January
30,
2009, P.B.
Liu, S., A. H. B. Duffy, R. I. Whitfield, I. M. Boyle, and I. McKenna, "Towards the Realization of
an Integrated Decision Support for Organizational Decision Making,"
International Journal of
Decision Support
Technology,
Vol. 1, No. 4, October-November, 2009, pp. 38-58.
Looney, CA. and A. M. Hardin, "Decision Support for Retirement Portfolio Management: Over-
coming Myopic Loss Aversion via Technology Design,"
Management Science,
Vol. 55, No. 10,
October, 2009, pp. 1688-1703.
Moore, N. C, "Game Utilizes Human Intuition to Help Computers Solve Complex Prob-
lems,"
University of Michigan News Service,
July 27, 2009, available: http://www.ns.umich.
edu/htdocs/releases/story.php?id=7252.
National Research Council Committee on Technical and Privacy Dimensions of Information for
Terrorism Prevention and Other National Goals, "Protecting Individual Privacy in the Struggle
Against
Terrorists:
A Framework for Program Assessment," Washington,
DC:
National Academies
Press,
2008.
Nord, R., and E. Schmitz, "A Decision Support System for Personnel Allocation in the U.S. Army,"
in P. Gray (Ed.),
Decision Support and Executive Information Systems,
Englewood Cliffs, NJ:
Prentice-Hall, 1994, pp.
191-201.






SUGGESTED READINGS
Port, O., "Smart Programs Go to Work,"
Business
Week,
March 2, 1995, pp. 97-99.
Rizakou, E. J., Rosenhead, and K. Reddington, "AIDSPLAN: A Decision Support Model for
Planning the Provision of HIV/AIDS Related Services,"
Interfaces,
Vol. 21, No. 3, 1991,
pp.
117-129.
Samuelson, D., "Road to the White House,"
OR/MS
Today,
Vol. 35, No. 5, October 2008.
Samuelson, D., "Unlocking the Door to the White House,"
ORMS Today
Vol. 23, No. 5, October,
1996.
Sauter, V L., "The Effect of 'Experience' upon Information Preferences,"
Omega: The International
Journal of Management Science,
Vol.
13,
No. 4, June, 1985a, pp. 277-284.
Sauter, V. L., "A Framework for Studying the Mergers of Public Organizations,"
Socioeconomic
Planning Sciences,
Vol. 19, No. 2, 1985b, pp. 137-144.
Sauter, V. L., and M. B. Mandell, "Using Decision Support Concepts to Increase the Utilization of
Social Science Information in Policy-Making,"
Evaluation and
Program
Planning,
Vol. 13, 1990,
pp.
349-358.
Sprague, R. H., and H. J. Watson,
Decision Support for Management,
Upper Saddle River, NJ:
Prentice-Hall, 1996.
Star, J., and J. Estes,
Geographic Information Systems,
Englewood Cliffs, NJ: Prentice-Hall, 1990.
Stovall Jr., A. A.,
Army
Command,
Leadership and Management: Theory and Practice,
Carlisle
Barracks, PA: U.S. Army War College, 1995-1996, pp. 17-18, 17-19, 18-2, and 18-3.
Sullivan, G., and K. Fordyce, "Decision Simulation: One Outcome of Combining Artificial In-
telligence and Decision Support Systems," in P. Gray (Ed.),
Decision Support and Executive
Information Systems,
Englewood Cliffs, NJ: Prentice-Hall, 1994, pp. 409^tl9.
Szeman, I., "Business Intelligence Past, Present and Future," SAS Institute, 2006, avail-
able:
http://www.sas.conVsearch/cs.html?url=http%3A//www.sas.corn/offices/europe/bulgaria/
downloads/saga_conf_sas.ppt&charset=iso-8859-l&qt=degree+of+intelligence+competitive
+
advantage+%2Bgraphic&col=extsas&n=l&la=en, viewed January 29, 2009.
Waterman, D. A.,
A Guide to Expert Systems,
Reading, MA: Addison-Wesley Publishing Company,
1986.
Watson, H. J. "Business Intelligence: Past, Present and Future,"
Communications of the
Association for Information Systems,
Vol. 25, No. 1, Article 39, 2009. Available at:
http://aisel.aisnet.org/cais/vol25/issl/39
Weiss, S., and N. Indurkhya,
Predictive Data Mining: A Practical Guide,
Burlington, MA: Morgan
Kaufmann, 1998.
Wishart, D.,
Whiskey
Classified:
Choosing Single Malts by Flavour,
London: Pavilion Books, 2006.
Wonnacott, T. H., and R. J. Wonnacott,
Introductory Statistics for Business and
Economics,
2nd ed.,
New York: Wiley, 1977.
QUESTIONS
1.
What is a model and why would a manager use one?
2.
Does a CASE tool use models? Describe them. Is it a DSS? If not, explain why it does
not have the attributes of a DSS. If so, explain how we might design CASE tools better
by considering DSS technology?
3.
Suppose you are developing a DSS to aid an MIS manager in deciding how to acquire
computers and computer components for her company. What kinds of models would
you provide in such a system? How would these models need to be integrated? What






MODEL MANAGEMENT SYSTEM
kinds of model management support do we need to facilitate model understandability
and/or sensitivity of a decision?
4.
What
are the
long-term implications for business when too much intelligence
is
included
inaDSS?
5.
How can a designer improve the users' understanding of results of a model in a DSS?
6. Suppose you were using a DSS to decide what courses to take for the next semester.
What kinds of models would you need? What kinds of sensitivity analyses would you
do?
7.
How can a designer ensure models in a DSS are integrated?
8. How can a DSS decrease a manager's anxiety about using models?
9. One of the primary things that differentiates
a DSS
from an MIS is that a
DSS
facilitates
analysis
of the data, whereas the MIS facilitates
reporting
of the data. Discuss the
difference between these two.
10.
There are multiple critical functions that a MBMS must provide, including alternative
generation, model selection, access to models, and sensitivity analysis. Discuss how
you might include these functions in a system that is intended to provide support for
someone selecting, a computer system.
11.
What are the long-term implications for business when too much intelligence
is
included
in a DSS?
12.
Describe three advantages of each of the kinds of modeling that we discussed in class.
13.
There are hundreds of DBMS packages on the market. Explain why there are no MBMS
packages on the market.
14.
What would be the advantages and disadvantages of using Monte Carlo simulation to
assess a DSS that provides advice about coursework and/or careers.
15.
J.S. Armstrong said, "Better predictions of how other parties will respond can lead to
better decisions." Discuss how you might build such a capability into a DSS.
16.
Malcoln Gladwell Published a book in 2005 called
Blink: The Power of Thinking
Without Thinking,
in which he claimed that frequently the intuitive, first unpression
decision (made in the first seconds) is a better decision than those supported by
signif-
icant analysis and data. Under what conditions do you believe this to be true? Defend
your position. If it is true (or when it is true), how would you provide decision support?
What are the implications for DSS if the author of the book
Blink
is correct in his
assessment of significant data.
17.
Discuss how Google's data mining and GapMinder's data analysis efforts could be
used to improve public policy discussion in the United States.
18.
How are models and analytics related? How are they different?
19.
What kinds of models do you use in your daily life?
20.
What attributes of a DSS make model use more attractive?
21.
Identify an article that appears in a newspaper or news magazine. What kinds of
models seem to be discussed in the article? Do the assumptions of the models seem
appropriate? What kinds of sensitivity testing did they discuss in the article? What
kinds of sensitivity testing do you think they should do?
22.
Suppose the problem for which you provided decision support required the decision
maker to utilize f-test to determine if part time employees were as productive as full






ON
THE WEB
195
time employees in a call center. Specifically, the decision maker compared the average
time on
a
call and the average number of calles that were handled, what specific decision
support would you provide to the decision maker.
ON
THE WEB
On the Web
for this chapter provides additional information about models, model base
management systems, and related tools. Links can provide access to demonstration pack-
ages,
general overview information, applications, software providers, tutorials, and more.
Additional discussion questions and new applications will also be added as they become
available.
•
Links provide access to information about model and model management products.
Links provide access to product information, product comparisons and reviews, and
general information about both models and the tools that support the models. Users
can try the models and determine the factors that facilitate and inhibit decision
making.
•
Links provide access to descriptions of applications and insights for
applications.
In
addition to information about the tools themselves, the Web provides links to world-
wide applications of those products. You can access chronicles of users' successes
and failures as well as innovative applications.
•
Links provide access to hints about how to use models.
These links provide real-
world insights into the use
and misuse
of models. These are descriptive and help
users to better formulate model management needs.
•
Links provide access to models
regarding
automobile purchase and
leasing.
Several
tools to help users purchase or lease an automobile are available on the Web. You
can scan links to determine what kinds of models are most useful under what
circumstances. Further, you can determine what kinds of impediments and what
kinds of model support are introduced by various modeling management tools.
Finally, the links can provide evaluations for model management capabilities.
You can access material for this chapter from the general web page for the book or directly
at http://www.umsl.edu/~sauterv/DSS4BI/mbms.html.





